Safeguard
AI Security•

OpenAI vs Anthropic vs Google vs Meta: Vulnerabilities Compared

OpenAI, Anthropic, Google, Meta and Microsoft compared on CVEs in their own AI products: counts, severity, weakness types, fix times and disclosure.

Safeguard Security Team
71 min read

By raw count, Microsoft has the most publicly disclosed CVEs in its own AI products: 75 on 16 November 2026, ahead of Amazon (53), Anthropic (43), Google (16), DeepSeek (15), OpenAI (14), Mistral AI (14), Meta (10) and Cohere (1). xAI has none. The open-source runtimes people use to serve open-weight models carry more than any vendor: vLLM alone has 140.

Those numbers say more about disclosure than about safety. The vendors at the top of the list run their own CVE Numbering Authorities, publish advisories for cloud services that needed no customer action, and ship open-source coding agents that researchers can read line by line. The vendors near the bottom either run almost everything server-side, or have most of their CVEs filed by third parties. This post puts the raw counts next to three normalised views, splits them by severity, weakness type and product layer, lists the notable individual cases with their fixes, and scores each vendor's disclosure programme. It is written for security and platform teams choosing, or already running, AI SDKs, coding agents, MCP connectors and inference servers. By the end you will know what the data shows for each vendor, why a raw count misleads, and what to measure for your own estate instead.

Contents

  • Key takeaways
  • Which LLM vendor has the most disclosed vulnerabilities?
  • How do the numbers change once normalised?
  • Why do raw CVE counts mislead?
  • How severe are the vulnerabilities each vendor discloses?
  • Which vulnerability types dominate each vendor?
  • Where in the AI application stack do these vulnerabilities live?
  • How fast is the count growing?
  • What are the most notable individual cases?
  • How fast do vendors fix reported vulnerabilities?
  • Which vendors run bug bounties, publish advisories and assign their own CVEs?
  • What does the data show for each vendor?
  • How do open-model runtimes compare?
  • OpenAI vs Anthropic, ChatGPT vs Claude, Gemini vs ChatGPT: how do the pairings compare?
  • Does a high count mean an insecure vendor?
  • What should a team building on these products measure instead?
  • Step by step: reproduce the numbers yourself
  • How we measured
  • Common mistakes when comparing AI vendors on security
  • Fix it without waiting for upstream
  • FAQ
  • Quick check
  • Sources

Key takeaways

Raw leader: Microsoft has 75 CVEs in its AI products, ahead of Amazon (53) and Anthropic (43), per CVE.org records read on 16 November 2026. Google has 16, DeepSeek 15, OpenAI and Mistral AI 14 each, Meta 10, Cohere 1 and xAI 0.

The MCP project adds 55 more. Its SDKs and reference servers sit inside Anthropic's CVE Numbering Authority scope, so we show them as a separate row rather than folding them into Anthropic's count or leaving them out.

Open-model runtimes outnumber every vendor: vLLM has 140 CVEs, llama.cpp 49 and Ollama 34.

Growth is recent. 185 of the 241 vendor CVEs (77%) were published in 2026, and 100 of them (41%) are in coding agents, CLIs and IDE extensions.

Disclosure channel separates the field more cleanly than the count does. Google (100%), Microsoft (99%), Amazon (98%), Anthropic (95%) and Meta (90%) published nearly all of their own CVEs through their own CNA or their own GitHub advisories. Every CVE for Mistral AI, DeepSeek and Cohere was filed by a third party.

Weakness mix follows product shape. Microsoft leads on prompt injection (15 CVEs, mostly in Microsoft 365 Copilot). Anthropic's profile is command injection (14) and path traversal (12) in Claude Code. Meta's is memory safety in ExecuTorch. vLLM's is resource exhaustion (62).

Several of the most serious events have no CVE at all: the DeepSeek database exposure, the March 2023 ChatGPT data exposure, the xAI API key on GitHub and the indexed Grok share links.

Counts are not risk. What matters is which of these products you run, whether the vulnerable path is reachable from your code, and how fast the fix reaches your build.

Bar chart of CVEs in each LLM vendor's own products: Microsoft 75, Amazon 53, Anthropic 43, Google 16, DeepSeek 15, OpenAI 14, Mistral AI 14, Meta 10, Cohere 1, xAI 0, MCP project 55, and runtimes vLLM 140, llama.cpp 49, Ollama 34

Figure 1. CVEs in each LLM vendor's own products, all years to 16 November 2026. Source: CVE.org records, OSV and the GitHub Advisory Database, queried 16 November 2026.

Which LLM vendor has the most disclosed vulnerabilities?

Microsoft has the most, with 75 CVEs in Copilot, GitHub Copilot, Azure AI and its agent frameworks. Amazon follows with 53 and Anthropic with 43. Google, DeepSeek, OpenAI and Mistral AI sit between 14 and 16, Meta has 10, Cohere 1 and xAI none, as of 16 November 2026.

We counted a CVE for a vendor only when the vendor's own product is the affected product on the authoritative CVE.org record, or when the advisory sits against a package the vendor publishes. Products built by third parties on top of a vendor's API, such as a ChatGPT wrapper or a Claude plugin someone else wrote, are excluded. Each CVE is counted once even when several packages carry it.

Table 1: CVEs by vendor, raw counts and severity

VendorCVEsCriticalHighMediumLowFirst CVE publishedAdvisories without a CVEMalicious release events
Microsoft75213420014 Feb 202300
Amazon535291903 May 202420
Anthropic4312810424 Jun 202510
Google16582110 Dec 202400
DeepSeek1534623 Sep 202500
OpenAI1426516 Jul 202400
Mistral AI1468005 Aug 202602
Meta10612123 Oct 202400
Cohere1001010 Oct 202600
xAI00000none00
MCP project (Anthropic CNA scope)5523119313 Jun 202560
vLLM (runtime)140743791117 Sep 202420
llama.cpp (runtime)4932915226 Feb 202400
Ollama (runtime)34423618 Apr 202400

Source: CVE.org records for every CVE, OSV and the GitHub Advisory Database for package advisories, all queried 16 November 2026. Severity uses the CNA's CVSS base score (v3.1 where given, otherwise v4.0, otherwise the ADP score on the same record). "Advisories without a CVE" are reviewed GitHub advisories that never received a CVE ID. The two Mistral AI events are the May 2026 compromised mistralai PyPI release and the compromised @mistralai npm releases, recorded as four advisory entries.

Read at face value, this table says Microsoft is the riskiest AI vendor and xAI the safest. Neither reading holds up, for reasons the next two sections make concrete.

Two scope notes matter for the top of the table. Nine of Microsoft's 75 CVEs are in Azure Machine Learning, a model-training and serving platform rather than a generative AI product; without them Microsoft has 66. And Amazon's count includes 11 CVEs in SageMaker, its equivalent platform. We kept both in, along with Google's Vertex AI, so the three cloud vendors are compared like for like.

How do the numbers change once normalised?

Normalised, the order reshuffles. In 2026 alone, Microsoft (55) and Amazon (49) still lead, but Mistral AI's 14 CVEs all arrived in under four months. Per affected product line, Mistral AI (14 in one product) and Amazon (8.8) lead. By share disclosed through the vendor's own channel, Google, Microsoft, Amazon and Anthropic lead.

There is no registry of "AI products" to divide by, the way the reference package studies divide advisories by the number of packages in npm or PyPI. So we used three denominators that each answer a different question, and we show all three rather than pick the flattering one.

Table 2: Normalised views

VendorCVEsPublished in 2026 to dateProduct lines with a CVECVEs per affected product lineShare published through the vendor's own channelMedian CVSSCritical or High
Microsoft7555107.599%8.173%
Amazon534968.898%7.364%
Anthropic433058.695%7.767%
Google161453.2100%8.581%
DeepSeek151327.50%6.347%
OpenAI14843.557%7.357%
Mistral AI1414114.00%8.8100%
Meta10125.090%9.870%
Cohere1111.00%4.80%
xAI000nonenonenonenone
MCP project55431not comparable91%7.360%

Source: as Table 1. "Own channel" means the CVE was assigned by the vendor's own CNA, or by GitHub's CNA from a security advisory the vendor published on its own repository.

Three bar panels: CVEs published in 2026 to date, CVEs per affected product line, and the share of CVEs published through the vendor's own channel, for each vendor

Figure 2. Three normalised views. Each panel has its own scale. Source: CVE.org, OSV and GitHub Advisory Database, queried 16 November 2026.

Per year: almost everything is recent

A per-year rate is the obvious normalisation, but it breaks down here because the histories are so short and so uneven. Mistral AI's first CVE was published on 5 August 2026; dividing 14 CVEs by a little over three months gives an annual rate near 50, which no one should quote. A shared window is fairer. In the 2026 calendar year to date, the order is Microsoft 55, Amazon 49, Anthropic 30, Google 14, Mistral AI 14, DeepSeek 13, OpenAI 8, Meta 1 and Cohere 1.

Meta is the outlier in the other direction. Nine of its ten CVEs were published in 2024 and 2025, almost all in ExecuTorch, and only one in 2026.

Per product line: concentration

The second view divides each vendor's CVEs by the number of distinct product lines that carry at least one. It measures concentration, not exposure, in the same way that advisories per vulnerable package showed how Maven's advisories cluster in a few large frameworks.

Mistral AI's 14 CVEs all sit in one product, the Mistral Vibe coding CLI. Anthropic's are concentrated too: 35 of 43 are in Claude Code, with the rest spread across the API client SDKs (4), Claude Desktop (2), the Claude Agent SDK (1) and the sandbox runtime (1). Microsoft's are spread across ten lines, led by Microsoft 365 Copilot (29) and GitHub Copilot (18). Google's 16 are spread across five, with the Agent Development Kit (6) and Vertex AI (5) the largest.

Concentration matters more to a security programme than the totals do. If your developers use Claude Code, Codex CLI or Mistral Vibe, most of that vendor's disclosed vulnerabilities land on one tool you can version-pin and update in one place.

Per disclosure channel: who told you

The third view asks who filed the CVE. A vendor that writes its own advisory, assigns its own CVE and publishes the fixed version is doing something different from a vendor whose vulnerabilities are catalogued by outside researchers and third-party CNAs.

Google (16 of 16), Microsoft (74 of 75), Amazon (52 of 53), Anthropic (41 of 43) and Meta (9 of 10) published nearly everything themselves. OpenAI sits in the middle at 8 of 14: its own CNA filed 7 and GitHub filed 1 from an OpenAI repository advisory, while MITRE, Google and JFrog filed the rest. Mistral AI (0 of 14), DeepSeek (0 of 15) and Cohere (0 of 1) had every CVE filed by a third party, mostly VulnCheck, HiddenLayer and VulDB.

None of this says the third-party CVEs are wrong. It says the vendor did not publish them, which matters to a team that wants to subscribe to one advisory feed per supplier.

Why do raw CVE counts mislead?

Raw counts mostly measure three things other than security: how much installable software a vendor ships, whether it runs its own CVE Numbering Authority, and whether it assigns CVEs to cloud-service bugs that customers never had to patch. A vendor with a closed, server-side product can have zero CVEs and still have serious incidents.

Four effects push the numbers around.

The first is installable surface. A CVE needs a product someone can identify and version. Coding agents, SDKs and desktop apps have version numbers and public repositories; a hosted chat service has neither. Anthropic, Amazon and Microsoft all ship open-source coding tools and agent SDKs on GitHub, where researchers can read every line. A vendor whose products are mostly hosted APIs gives researchers less to read and gives CVE assignment less to attach to.

The second is CNA status. Six of the ten vendors are CVE Numbering Authorities: OpenAI, Anthropic, Google (twice, as Google and Google Cloud), Microsoft, Meta and Amazon. A CNA can assign CVEs as part of its own release process, which lowers the cost of disclosure to near zero. Both OpenAI and Anthropic are also part of the CVE program's Frontier AI Researcher CNA pilot, which lets them assign CVEs for vulnerabilities they discover in other people's software.

The third is policy on cloud CVEs. In June 2024 Microsoft said it would issue CVEs for critical cloud-service vulnerabilities even when customers have nothing to install, and tag them as exclusively hosted. That decision accounts for much of its count: 55 of Microsoft's 75 CVEs are in hosted services and apps, including EchoLeak. Most other vendors fix server-side issues silently or credit the researcher without a CVE.

The fourth is research attention. Coding agents became the most-studied AI attack surface in 2025, and the vendors with the most popular agents drew the most reports. Claude Code, Codex CLI, Gemini CLI, GitHub Copilot, Kiro and Mistral Vibe together account for most of the 2025 and 2026 growth.

The consequence is easy to state. A higher count often reflects a more mature disclosure programme and a larger open surface, not a weaker product. A count of zero, as xAI has, means only that nothing has been filed.

How severe are the vulnerabilities each vendor discloses?

Most disclosed vulnerabilities in LLM vendor products are rated High. Google (81% Critical or High), Microsoft (73%) and Meta (70%) carry the most severe mixes, while Anthropic (67%), Amazon (64%) and OpenAI (57%) sit close behind. Severity tracks product type: local code execution in coding agents and cloud information disclosure both score high.

100 percent stacked bars showing the share of Critical, High, Medium and Low CVEs for each vendor, the MCP project and the three runtimes

Figure 3. Severity mix per vendor by CVSS base score band. Source: CNA scores on CVE.org records, queried 16 November 2026.

A few patterns stand out.

Microsoft has 21 Critical CVEs; the other nine vendors have 28 between them. Most of Microsoft's are cloud-service issues in Microsoft 365 Copilot and Copilot Studio that Microsoft scored itself, often 9.3 or higher because they allow information disclosure across a network without authentication. EchoLeak (CVE-2025-32711) is the best known, at 9.3.

Anthropic's profile is the narrowest: one Critical (the Claude Agent SDK argument injection, CVE-2026-96620, at 9.2), 28 High, 10 Medium and 4 Low. The High cluster is almost entirely Claude Code approval and sandbox bypasses, which GitHub's CNA scored consistently at 7.7 to 8.7 on CVSS v4.0.

Meta's 60% Critical share comes from six ExecuTorch memory-safety CVEs, each scored 9.8 in the ADP data on CVE.org. Meta's own CNA score for the best-known Meta case, the Llama Stack pickle flaw CVE-2024-50050, is 6.3. Independent analysts rated the same issue 9.3. Both numbers are defensible; they assume different deployment exposure.

Mistral AI has no Medium or Low CVEs at all. All 14 are Critical or High, scored by the third-party CNAs that filed them.

Scores are not calibrated across CNAs. Microsoft scores its cloud CVEs on CVSS v3.1, while GitHub's CNA, which filed most of Anthropic's and the MCP project's, uses v4.0. Treat the bands as a rough guide.

Which vulnerability types dominate each vendor?

Command and code injection (46 CVEs) and authentication or access-control flaws (48) are the most common types across the ten vendors, followed by path traversal (24), prompt injection (19) and SSRF or origin issues (17). Each vendor has a signature class that reflects what it ships, not how carefully it codes.

Table 3: Vulnerability types by vendor

VendorCommand and code injectionPath traversalAuth and access controlPrompt injectionSSRF, origin, redirectDeserializationMemory safetyResource exhaustionInfo exposure, permissionsInput validation, other
Microsoft2031515710059
Amazon831145100615
Anthropic141260300035
Google0130200037
DeepSeek1140000027
OpenAI3120000017
Mistral AI0270000014
Meta0000007012
Cohere0100000000
MCP project2514010001446
vLLM5150791262732
llama.cpp10000136803
Ollama1640409415

Source: CWE IDs on each CVE.org record, grouped into families by Safeguard. A CVE with several CWEs is counted once, in the first matching family in the column order above (prompt injection first). CVEs with no CWE fall in "other".

Heatmap of 13 organisations by 10 weakness families, with cell counts shaded by share of each vendor's total

Figure 4. Vulnerability types per vendor. Shade shows each family's share of that vendor's total so small and large vendors compare fairly. Source: CWE IDs on CVE.org records, queried 16 November 2026.

What the weakness mix says about each vendor

Microsoft is the only vendor with a large prompt injection column, and that is a disclosure choice as much as a product fact. Fifteen of its CVEs are injection flaws in Copilot assistants that let an attacker disclose information, the class EchoLeak made famous. Microsoft labels several of them "AI command injection" on the CVE record. Other vendors with hosted assistants have had comparable prompt injection reports (ShadowLeak against ChatGPT, the Gemini Trifecta against Gemini), fixed them, and not assigned CVEs.

Anthropic's mix is command injection (14) and path traversal or link following (12). Almost all of it is Claude Code: a sed, find, echo or rg invocation that slips past the read-only allowlist, a symlink or zsh clobber that writes outside the workspace, a repository settings file that runs before the trust dialog. The same shape appears in OpenAI's Codex CVEs, Google's Gemini CLI CVEs, GitHub Copilot CLI, Amazon's Kiro and Mistral Vibe. It is the coding-agent class, and every vendor with a coding agent has it.

Amazon and Mistral AI lead on authentication and access control relative to their size. For Amazon that is mostly its open-source MCP servers and Strands Agents tools; for Mistral AI it is Mistral Vibe's command permission gate.

Meta's profile is memory safety, from ExecuTorch, its on-device inference runtime. That is the one vendor profile that looks like the runtimes at the bottom of the table rather than like the other vendors.

Deserialization barely appears among the vendors (two CVEs), even though the best-known Meta case is a pickle flaw. Meta's record for CVE-2024-50050 carries no CWE, so it falls in "other". In vLLM, deserialization is a real cluster, with nine CVEs.

The MCP project's mix is a web-security mix: authentication and session binding (14), resource exhaustion in the HTTP transports (14), and SSRF, DNS rebinding and redirect issues (10).

Where in the AI application stack do these vulnerabilities live?

Most disclosed vulnerabilities in LLM vendor products sit in the software around the model, not in the model. The busiest layers are coding agents and CLIs, agent frameworks and SDKs, the MCP connector layer, and self-hosted inference servers. Hosted chat services produce fewer CVEs but the largest data exposures.

The AI application stack in six layers, from hosted model APIs to developer machines, with the vulnerability classes and example CVEs found at each layer

Figure 7. Where vulnerabilities live in an application built on LLM vendor products. Layers are ordered from the vendor's cloud down to the developer's laptop. Source: Safeguard analysis of CVE.org, GitHub Advisory Database and OSV records, as of 16 November 2026.

Read from the top down, the stack looks like this.

Hosted model services and chat apps. ChatGPT, Claude.ai, Gemini, Microsoft 365 Copilot, Meta AI, Grok, Le Chat and DeepSeek's chat service. The vendor runs the code, so there is nothing for you to patch. When something goes wrong here it tends to be a data exposure (the March 2023 ChatGPT Redis bug, the DeepSeek ClickHouse database, the Meta AI prompt identifier bug) or an indirect prompt injection that turns the assistant into an exfiltration channel (EchoLeak, ShadowLeak, the Gemini Trifecta). CVEs are the exception rather than the rule at this layer. Microsoft assigns them for cloud services; most other vendors fix server-side issues silently or credit researchers without a CVE.

Connectors and MCP. The layer that lets a model reach mail, files, repositories and internal tools. Its vulnerabilities look like classic web and network bugs because that is what they are: missing authentication on a local proxy, DNS rebinding, path validation by string prefix, OAuth tokens accepted for the wrong audience, cross-origin redirects that leak headers. The MCP SDKs and reference servers alone had 55 CVEs and 6 further advisories by 16 November 2026, and Amazon's own MCP servers and gateway added 14 more.

Agent frameworks and SDKs. Semantic Kernel, Google ADK, Llama Stack, Strands Agents, Bedrock AgentCore, the Claude Agent SDK, and the API client libraries. Here the classic classes return in AI clothing: deserialization of untrusted data, code injection through template or filter evaluation, argument injection into a subprocess, insecure default file permissions for agent memory.

Coding agents and CLIs. Claude Code, Codex CLI, Gemini CLI, GitHub Copilot CLI, the Copilot extensions in Visual Studio and VS Code, and Amazon Q Developer. This is the fastest-growing layer. The pattern is consistent across vendors: an allowlist or approval prompt that a crafted command, symlink, git setting or workspace file can step around, so that content in a repository the agent reads becomes a command the agent runs.

Inference servers and on-device runtimes. vLLM, Ollama, llama.cpp and ExecuTorch. These parse untrusted binary input (model files, tensors, images, video) and often listen on a network port. The result is a memory safety and resource exhaustion profile closer to a media codec than to a web app, plus the pickle problem wherever Python processes exchange objects.

Model artifacts. Pickle-based checkpoints execute code on load by design. GGUF, safetensors and ONNX were built to avoid that, but their parsers still have bugs, as the Ollama GGUF out-of-bounds read (CVE-2026-7482) and the ExecuTorch overflows show.

The practical consequence is that a team's exposure depends far more on which of these layers it deploys than on which vendor it picked. A company that only calls a hosted API through an official SDK carries a small fraction of the risk of a company that self-hosts vLLM, gives coding agents shell access on developer laptops, and wires both into internal systems through MCP.

Which product layer each vendor's CVEs land in

Splitting each vendor's count by stack layer shows why the vendors look so different. Across the ten vendors, coding agents, CLIs and IDE extensions account for 100 of 241 CVEs (41%), hosted services and apps for 74 (31%), SDKs and agent frameworks for 45 (19%), MCP servers for 14 (all Amazon's) and inference runtimes for 8 (Meta's ExecuTorch and Microsoft's DeepSpeed).

Stacked bars splitting each vendor's CVEs across hosted services, coding agents, SDKs and frameworks, MCP and connectors, and inference runtimes

Figure 10. Which product layer each vendor's CVEs land in. Layer assigned from the affected product on each CVE.org record. Source: CVE.org, OSV and GitHub Advisory Database, queried 16 November 2026.

Microsoft's count is mostly hosted services (55 of 75). Anthropic's is mostly a coding agent (35 of 43). Amazon's is spread across SDKs and agent frameworks (24), MCP servers (14) and coding tools (14, mostly Kiro). Google's is split between SDKs (8, mainly the Agent Development Kit) and hosted services (6, mainly Vertex AI). OpenAI's is mostly Codex (10 of 14, across the CLI and the desktop app). Mistral AI's and most of DeepSeek's are coding agents too. Meta's is an inference runtime.

So when someone asks "is OpenAI more secure than Anthropic?", the honest first answer is that the two vendors' CVEs mostly describe coding agents, and the question becomes whether Codex CLI or Claude Code has had more disclosed vulnerabilities. Claude Code has: 35 against Codex's 10. That also reflects Claude Code's earlier public release, its open GitHub advisory feed and the volume of research it attracted.

How fast is the count growing?

Very fast, and almost all of it is recent. Of the 241 vendor CVEs, 3 were published in 2023, 9 in 2024, 44 in 2025 and 185 in 2026 to 16 November. The inflection follows the public release of coding agents and agent SDKs in mid 2025, and the arrival of third-party CNAs that file AI-tool CVEs at volume.

Cumulative CVE count by month for Microsoft, Amazon, Anthropic, Google and OpenAI, with other vendors in grey, January 2023 to November 2026

Figure 9. Cumulative CVEs by vendor, 2023 to 2026. Source: CVE.org publication dates, queried 16 November 2026.

Amazon's curve is the steepest in 2026: 49 of its 53 CVEs were published this year, as Kiro, Strands Agents, Bedrock AgentCore and the AWS MCP servers drew research. Microsoft's climbs steadily from mid 2024, when it began assigning cloud-service CVEs. Anthropic's starts in June 2025 and rises with each Claude Code release cycle. Mistral AI's and DeepSeek's lines barely exist before August 2026, then jump as VulnCheck, HiddenLayer and VulDB filed batches against Mistral Vibe and DeepSeek Harness.

Two cautions. CVE publication dates lag the fix, sometimes by months, so the curves show when the world learned about a vulnerability, not when it existed. And a batch from one researcher or one CNA can add ten CVEs in a day, which is visible as the vertical steps in Figure 9.

What are the most notable individual cases?

The most consequential cases split into two kinds: CVEs in installable tools that users had to update (Claude Code, Codex CLI, Gemini CLI, Amazon Q, Ollama, Llama Stack) and server-side exposures with no CVE at all (ChatGPT in 2023, DeepSeek's database, Meta AI's prompt bug, xAI's leaked key). Every vendor in this comparison appears at least once.

Table 4: Notable cases, verified

DateVendor and productCVEWhat happenedFix
20 Mar 2023OpenAI, ChatGPTNone for ChatGPT (root cause in redis-py, CVE-2023-28858)A Redis client bug let some users see other users' chat titles, and exposed partial payment details of about 1.2% of ChatGPT Plus subscribers active in a nine-hour windowOpenAI patched the bug, added checks and notified affected users
5 May 2024 (reported)OllamaCVE-2024-37032, "Probllama"Path traversal through an unvalidated model digest allowed file write and remote code execution (Wiz)Fix committed within about four hours; 0.1.34 released 8 May 2024
5 Jul 2024OpenAI, ChatGPT for macOSCVE-2024-40594The app opted out of the sandbox and stored conversations in cleartext readable by other appsUpdate encrypting local conversations
6 Aug 2024Microsoft, Copilot StudioCVE-2024-38206 (CVSS 8.5)SSRF protection bypass let an authenticated user reach internal service endpoints (Tenable)Server-side fix, no customer action
10 Oct 2024 (fixed)Meta, Llama StackCVE-2024-50050 (CNA 6.3)Pickle used for socket communication over pyzmq allowed remote code execution (Oligo Security)0.0.41 replaced pickle with JSON
24 Jan 2025 (fixed)Meta, Meta AINoneGuessable prompt-edit IDs let a logged-in user read other users' prompts and AI responses (Sandeep Hodkasia)Fixed server-side; Meta paid a $10,000 bounty and found no evidence of abuse
29 Jan 2025DeepSeekNoneA publicly accessible ClickHouse database exposed over a million log lines including chat history, secret keys and backend details (Wiz Research)DeepSeek secured it promptly after Wiz's report
1 May 2025 (reported)xAINoneAn xAI employee's API key on GitHub gave access to at least 60 private and fine-tuned models; alerted 2 March, still valid 30 April (GitGuardian, KrebsOnSecurity)Repository removed from GitHub hours after escalation to xAI
11 Jun 2025Microsoft, Microsoft 365 CopilotCVE-2025-32711, "EchoLeak" (CVSS 9.3)Zero-click indirect prompt injection through email let an attacker exfiltrate data from Copilot's context (Aim Labs)Server-side fix, no customer action
13 Jun 2025MCP project, MCP InspectorCVE-2025-49596 (CVSS 9.4)No authentication between the Inspector client and proxy allowed remote code execution from a malicious web page (Oligo Security)0.14.1
24 Jun 2025Anthropic, Claude Code IDE extensionsCVE-2025-52882 (CVSS 8.8)The extensions accepted WebSocket connections from arbitrary origins1.0.24
23 Jul 2025Amazon, Q Developer for VS CodeCVE-2025-8217Injected code shipped in extension 1.84.0 through an over-scoped GitHub token in the build configuration; a syntax error stopped it running1.85.0, credentials revoked (AWS-2025-015)
25 Jul 2025 (fixed)Google, Gemini CLINonePrompt injection in a context file plus an allowlist bypass allowed silent command execution (Tracebit)0.1.14
5 Aug 2025Anthropic, Claude CodeCVE-2025-54794 (7.7), CVE-2025-54795 (8.7)Path restriction bypass by prefix matching; command injection that bypassed the approval prompt (Cymulate, "InversePrompt")0.2.111 and 1.0.20
7 Aug 2025Meta, ExecuTorchCVE-2025-30405 and five related CVEsInteger and heap overflows when loading models, potentially leading to code executionFixed in ExecuTorch commits named on each record
Early Aug 2025 (fixed)OpenAI, ChatGPT Deep ResearchNone, "ShadowLeak"Zero-click prompt injection in an email made the agent exfiltrate inbox data from OpenAI's cloud (Radware)Fixed server-side; disclosed 18 Sep 2025
12 Aug 2025Microsoft, GitHub Copilot and Visual StudioCVE-2025-53773 (CVSS 7.8)Prompt injection leading to local code executionVisual Studio 2022 17.14.12
20 Aug 2025xAI, GrokNoneShared conversation links were indexed by search engines; Forbes counted more than 370,000No public statement from xAI found
19 Sep 2025 (fix verified)Anthropic, Claude Desktop extensionsNone foundCommand injection in the official Chrome, iMessage and Apple Notes extensions (Koi Security, CVSS 8.9 per Koi)Claude Desktop 0.1.9
22 Sep 2025OpenAI, Codex CLICVE-2025-59532 (CVSS 8.6)A model-generated working directory could be treated as the sandbox's writable root0.39.0
30 Sep 2025Google, Gemini Cloud Assist, Search and browsingNone, "Gemini Trifecta"Three prompt injection paths to private data exfiltration (Tenable)Remediated by Google
Nov 2025OpenAI, API platform via MixpanelNoneA breach at analytics vendor Mixpanel exposed names, emails and approximate locations of some API users; no chats, keys or payment dataOpenAI ended its use of Mixpanel
6 Feb 2026Microsoft, Semantic Kernel .NETCVE-2026-25592 (CVSS 10.0)Arbitrary file write through the SessionsPythonPlugin during AI agent function calling1.71.0
13 Apr 2026Google, Agent Development KitCVE-2026-4810 (CVSS 9.3)Code injection and missing authentication allowed unauthenticated remote code executionReleases after 1.28.1
14 Apr 2026OpenAI, Codex CLICVE-2025-61260 (CVSS 9.8)A malicious MCP configuration file in a repository could execute code when a user ran codex inside itReleases after 0.23.0
11 to 18 May 2026Mistral AI, Python and TypeScript SDKsNone (GHSA-wx9m-wx4f-4cmg, GHSA-jgg6-4rpr-wfh7)A malicious mistralai 2.4.6 was uploaded to PyPI outside the release pipeline, and worm-driven compromised @mistralai npm versions were publishedVersions removed; PyPI project quarantined; Mistral reported no infrastructure compromise
24 Jun 2026Google, Gemini CLICVE-2026-12537 (CVSS 10.0)OS command injection in the container launcher on headless CI allowed pre-sandbox code execution0.39.1; run-gemini-cli action 0.1.22
1 Sep 2026OpenAI, Codex DesktopCVE-2026-19593 (CVSS 9.8)Opening a workspace with attacker-controlled Git configuration could make Git run an attacker's filterFixed builds named on the record
16 Oct 2026Anthropic, Claude Agent SDK for PythonCVE-2026-96620 (CVSS 9.2)The resume option was passed to the Claude CLI unsanitised, allowing argument injection and command execution0.2.121
27 Oct 2026Mistral AI, Mistral VibeCVE-2026-116843 (CVSS 8.8)Remote code execution through a permission check bypass in the CLI3.0.3

Sources for every row are listed in the Sources section. CVSS values are the CNA scores on the CVE.org record.

Timeline of 23 notable LLM vendor CVEs and incidents from March 2023 to October 2026

Figure 5. Timeline of notable vulnerabilities and incidents, 2023 to 2026. Source: CVE.org records, vendor advisories and researcher reports, as of 16 November 2026.

Three things in the table are worth drawing out.

The cases without CVEs are some of the largest. DeepSeek's exposed database, the ChatGPT Redis bug, Meta AI's prompt identifier bug and the Grok share links all exposed other users' data. None has a CVE, because CVEs describe flaws in identifiable products, and a misconfigured server or a share-link design does not fit that model neatly. Any comparison that only counts CVEs will miss them.

Supply chain compromise is now on the list. The Amazon Q extension in July 2025 and the Mistral AI SDK releases in May 2026 were not coding mistakes. In both, an attacker got code into a release channel the vendor controlled. In both, the malicious code was caught quickly and the vendor published an advisory. These are the cases where install-time checks matter more than CVE matching, because there is no vulnerable version range to match until after the fact.

The coding-agent pattern repeats across vendors. Claude Code, Codex CLI, Gemini CLI, GitHub Copilot, Kiro and Mistral Vibe all had CVEs where content inside a repository, such as a settings file, a Git config, an MCP config or a crafted command, ran before or around the user's approval. If your developers run any of these tools, that is the class to plan for.

How fast do vendors fix reported vulnerabilities?

Where both dates are public, documented fixes took from three days (Ollama's Probllama) to roughly four months (EchoLeak). Most fell between two and seven weeks. There are too few cases with published report and fix dates to rank vendors on speed, so treat these as examples rather than a league table.

Dot chart of days from private report to fix: Ollama 3 days, Llama Stack about 16, Gemini CLI 28, Meta AI 29, ShadowLeak about 48, EchoLeak about 4 months

Figure 11. Time from private report to fix, documented cases only. Source: Wiz, Oligo Security and Meta, Tracebit, TechCrunch, Radware and secondary reporting on Aim Labs, as of 16 November 2026.

Ollama committed a fix for Probllama about four hours after Wiz reported it on 5 May 2024, and released 0.1.34 three days later. Meta's Llama Stack fix took about two weeks from Oligo's late-September report to the 0.0.41 release on 10 October 2024. Google fixed the Tracebit Gemini CLI issue in 28 days, after first triaging it as a lower priority and then reclassifying it as P1, S1. Meta fixed its prompt identifier bug in 29 days. OpenAI fixed ShadowLeak about seven weeks after Radware's report. Microsoft's EchoLeak fix took roughly four months from Aim Labs' January 2025 report, according to secondary reporting; neither Microsoft nor Aim Labs' public pages we could retrieve gave both dates.

The time that matters for your team is a different interval. It runs from a fixed version being published to that version running on every developer laptop, CI runner and inference server you operate. For hosted services the vendor closes that gap alone. For everything installable, it is yours.

Which vendors run bug bounties, publish advisories and assign their own CVEs?

Six vendors score on all five disclosure signals we checked: OpenAI, Anthropic, Google, Microsoft, Meta and Amazon each run a public bug bounty, publish a disclosure policy and advisories, are CVE Numbering Authorities and serve a security.txt. Mistral AI, xAI and Cohere score on some. DeepSeek scored on none we could find.

Table 5: Disclosure maturity scorecard

OrganisationPublic bug bountyDisclosure policyPublishes advisoriesCVE Numbering Authoritysecurity.txt
OpenAIYes, BugcrowdYes, coordinated disclosure and CVE assignment policiesYes, trust portal and GitHub advisoriesYes, desktop apps, mobile apps and SDKsYes
AnthropicYes, HackerOne, public since May 2026YesYes, GitHub advisories and trust portalYes, including MCP SDKs and reference serversYes
GoogleYes, Bug Hunters and the AI VRP launched October 2025YesYes, Google Cloud security bulletins and GitHub advisoriesYes, as Google and as Google CloudYes
MicrosoftYes, MSRC bounties including CopilotYesYes, Security Update GuideYesYes
MetaYes, Meta Bug BountyYesYesYesYes
AmazonYes, Amazon VRP on HackerOneYes, AWS vulnerability reportingYes, AWS security bulletinsYesYes
Mistral AIReport form on HackerOne; no public bounty foundLinked from security.txtOne GitHub advisory set (May 2026)NoYes
xAIYes, HackerOneStated in security.txtNone foundNoYes
CohereReferenced in its Responsible Disclosure PolicyYes, Trust CenterNone foundNoNo
DeepSeekNone foundNone foundNone foundNoNo
OllamaNoYes, SECURITY.md via GitHub private reportingNo GitHub advisories; CVEs filed by othersNoNo
vLLMNoYes, SECURITY.md and a vulnerability management teamYes, more than 100 GitHub advisoriesNoNo

Checked 16 November 2026. Sources: each domain's /.well-known/security.txt, the CVE program's published CNA list, vendor policy pages and GitHub repository advisories. "None found" means we searched and could not find one, not that one cannot exist.

Scorecard grid of 12 organisations by five disclosure signals, each marked yes, partial or no

Figure 6. Disclosure maturity scorecard, checked 16 November 2026. Source: vendor security.txt files, the CVE program CNA list, vendor policy pages and GitHub repository advisories.

The scorecard and the CVE count tell the same story from two directions. The six vendors that score fully are the six with their own CNA, and five of them published 90% or more of their CVEs themselves. OpenAI's lower own-channel share (57%) comes from CVEs MITRE and others assigned before or outside its CNA scope, which covers installed software rather than hosted ChatGPT.

The tier below is uneven. xAI runs a HackerOne bug bounty and serves a security.txt, but publishes no advisories and has no CVEs. Mistral AI's security.txt points to a HackerOne submission form, and its only published advisories are the two from the May 2026 package compromise; every Mistral Vibe CVE came from outside. Cohere's Trust Center hosts a Responsible Disclosure Policy that references a bug bounty, but it has no security.txt and its one CVE was filed by VulnCheck. Wiz reported in January 2025 that DeepSeek had no established security contact when it found the exposed database, and we found none in November 2026 either.

What does the data show for each vendor?

Each vendor's record reflects what it ships. Microsoft and Amazon disclose the most because they ship the most AI products and assign their own CVEs. Anthropic's record is dominated by one open-source coding agent. Google, OpenAI and Meta disclose less, in fewer products. Mistral AI, DeepSeek, Cohere and xAI have thin or third-party records.

The vendors appear in the order of Table 1. Each paragraph says what the data shows and what it does not.

Microsoft: 75 CVEs, most in hosted Copilot services

Microsoft's 75 CVEs span ten product lines: Microsoft 365 Copilot (29), GitHub Copilot across VS Code, Visual Studio, JetBrains and the CLI (18), Azure Machine Learning (9), Copilot Studio (5), consumer Copilot and Copilot in Edge (5), Azure AI Foundry (4), Semantic Kernel (2), and one each in DeepSpeed, Prompt Flow and Copilot in Azure. Microsoft's own CNA assigned 70 of them; GitHub assigned four more from Microsoft's own repositories.

The count is high largely because Microsoft decided in June 2024 to publish CVEs for critical cloud-service vulnerabilities even when customers had nothing to install. EchoLeak, Copilot Studio's SSRF and most of the 15 prompt injection CVEs are in that category. For a buyer, that policy is useful: it creates a public, dated record of what was fixed in a service you cannot inspect. It also means Microsoft's count is not comparable with vendors that fix hosted issues silently.

Amazon: 53 CVEs across agents, SDKs and MCP servers

Amazon's 53 CVEs are spread across AWS MCP servers and the MCP Gateway (14), Kiro (12), SageMaker (11), Bedrock AgentCore (8), Strands Agents (7) and one in the Amazon Q Developer extension. Its own CNA assigned 50. Forty-nine were published in 2026, the steepest rise of any vendor, as Kiro, AgentCore and Strands moved from launch to wide use.

The Amazon case most teams will remember has a modest CVE: CVE-2025-8217, the injected code in Amazon Q Developer for VS Code 1.84.0. AWS's bulletin explains that an over-scoped GitHub token in the build configuration let an outsider commit code that shipped in the extension, and that a syntax error stopped it running. AWS revoked the credentials and released 1.85.0.

Anthropic: 43 CVEs, 35 of them in Claude Code

Anthropic's 43 CVEs are concentrated: Claude Code (35), the Python and TypeScript API SDKs (4, all in the local memory tool), Claude Desktop (2), the Claude Agent SDK (1) and the sandbox runtime (1). Anthropic published 41 of the 43 itself, nearly all as GitHub security advisories on its own repositories, which GitHub's CNA turned into CVEs. Anthropic became a CNA in its own right, with a scope that also covers the MCP SDKs and reference servers.

The Claude Code CVEs follow one pattern. A command the read-only allowlist treats as safe (echo, sed, find, rg, a cd into a crafted path), a symlink, a git worktree or a repository settings file lets content in the working directory run or write before or around the user's approval. Severity is consistently High (7.7 to 8.7 on CVSS v4.0) and fixes ship in the next Claude Code release. Anthropic's single Critical is CVE-2026-96620 in the Claude Agent SDK for Python, fixed in 0.2.121.

The MCP project's 55 CVEs are counted separately in this post. If you add them, Anthropic's total within its CNA scope is 98. If you consider MCP a community project, as its governance now is, Anthropic's own products account for 43.

Google: 16 CVEs, all self-published, most in agent tooling

Google's 16 CVEs sit in the Agent Development Kit (6), Vertex AI (5), Gemini CLI (2), Gemini Enterprise Agent Platform (2) and the Gemini app (1). Google published all 16 itself, through Google's and Google Cloud's two CNAs. Five are Critical, including CVE-2026-12537 in Gemini CLI's container launcher (CVSS 10.0 on v4.0, fixed in 0.39.1) and CVE-2026-4810 in the Agent Development Kit (9.3).

Google's record also shows the gap between fixes and CVEs. Tracebit's July 2025 Gemini CLI report and Tenable's Gemini Trifecta were both fixed without a CVE. Google's AI Vulnerability Reward Program, launched in October 2025, now pays for exactly this class of AI product issue.

DeepSeek: 15 CVEs, all filed by others, plus a major data exposure

DeepSeek's 15 CVEs are in DeepSeek Harness, its open-source agent harness (12), and in DeepSeek's models and chat service (3, all assigned by MITRE, including two cross-site scripting issues through model-generated content). None was published by DeepSeek. VulDB filed eight, MITRE five and VulnCheck two. Several MITRE records carry thin descriptions, and one notes that third parties dispute it.

The more significant DeepSeek event has no CVE. On 29 January 2025, Wiz Research found a publicly accessible ClickHouse database belonging to DeepSeek that exposed more than a million log lines, including chat history, secret keys and backend details. DeepSeek secured it promptly after Wiz's report.

OpenAI: 14 CVEs, most in Codex

OpenAI's 14 CVEs are in Codex CLI and the Codex desktop apps (10), the ChatGPT apps (2), the Atlas browser (1) and Operator (1). OpenAI's own CNA assigned seven, all in 2026, after it became a CNA with a scope covering its desktop apps, mobile apps and SDKs. MITRE, Google, JFrog and GitHub assigned the rest.

OpenAI's most widely reported security event remains the March 2023 ChatGPT exposure, caused by a bug in the open-source redis-py client. OpenAI published a detailed account at the time, which set an early standard for incident transparency in this market. Its later cases, ShadowLeak in ChatGPT Deep Research and the Mixpanel vendor breach in November 2025, were both disclosed publicly, the first by the researcher and the second by OpenAI.

Mistral AI: 14 CVEs in one product, plus a package compromise

All 14 Mistral AI CVEs are in Mistral Vibe, its coding CLI, and all were published between August and November 2026 by VulnCheck (8) and HiddenLayer (6). Every one is High or Critical, mostly command permission bypasses of the same kind seen in other coding agents. Mistral AI published none of them itself.

In May 2026 an attacker uploaded a malicious mistralai 2.4.6 to PyPI outside Mistral AI's release pipeline, and a worm linked to a separate npm incident published compromised @mistralai npm versions through an affected developer device. Mistral AI published GitHub advisories for both, removed the versions and reported no compromise of its infrastructure. These are not CVEs, but any team that installed those exact versions was exposed.

Meta: 10 CVEs, mostly ExecuTorch memory safety

Meta's 10 CVEs are in ExecuTorch, its on-device inference runtime (7), and Llama Stack (3). Meta's own CNA assigned nine. Eight of the ten were published in 2025, and only one in 2026.

The ExecuTorch CVEs are integer and heap overflows in model loading, the classic risk of parsing an untrusted binary file. The Llama Stack pickle flaw, CVE-2024-50050, is the better-known case. It also shows how severity can differ: Meta's CNA scored it 6.3, while Oligo Security and Snyk rated it 9.3. Separately, Meta fixed a Meta AI bug in January 2025 that let users read other users' prompts and responses, and paid the reporter a $10,000 bounty.

Cohere: 1 CVE

Cohere has one CVE, a tar-slip path traversal in the Cohere Python SDK filed by VulnCheck in October 2026. Cohere's products are mostly enterprise APIs and private deployments, so the public surface for researchers is small. Its Trust Center publishes a Responsible Disclosure Policy that references a bug bounty programme.

xAI: 0 CVEs, two documented incidents

xAI has no CVEs in any product. It does have two documented incidents. In 2025 an xAI employee's API key sat in a public GitHub repository for about two months with access to at least 60 private and fine-tuned models, until GitGuardian escalated to xAI and the repository was removed. In August 2025, Forbes reported that more than 370,000 Grok conversations shared through Grok's share feature had been indexed by search engines. xAI runs a HackerOne bug bounty and publishes a security.txt, but we found no published security advisories.

Zero CVEs here should be read as "nothing filed," not "nothing found." xAI ships few installable products, and its hosted services would only appear in the CVE record if xAI or a researcher chose to request an ID.

How do open-model runtimes compare?

The three most popular open-model runtimes have more CVEs than any vendor: vLLM 140, llama.cpp 49 and Ollama 34, as of 16 November 2026. They are open-source projects that parse untrusted model files and requests, often on a network port, so their profiles look like media codecs and web servers rather than AI assistants.

vLLM's 140 CVEs are dominated by resource exhaustion (62), followed by memory safety (12), deserialization (9) and SSRF (7). 113 were published in 2026. Its deserialization cluster includes the pickle-based remote code execution flaws in distributed inference transports: CVE-2025-47277 in the PyNcclPipe KV cache transfer (fixed in 0.8.5) and CVE-2025-32444 in the Mooncake integration. GitHub's CNA filed 71 from vLLM's own advisories, and VulnCheck filed 49. vLLM publishes more than 100 GitHub security advisories and has a named vulnerability management team.

llama.cpp's 49 CVEs are mostly memory safety (36), from GGUF parsing, tokenisation and the server. VulnCheck, GitHub, VulDB and Cisco Talos filed most of them.

Ollama's 34 CVEs mix memory safety in the GGUF loader (9), path traversal (6) and missing authentication. CVE-2025-63389 records that Ollama's API endpoints required no authentication through 0.12.3. Ollama publishes a SECURITY.md that routes reports to GitHub private reporting, but every one of its CVEs was filed by someone else, most often MITRE, Protect AI's huntr, CERT Polska, VulnCheck and VulDB. Ollama's response to Wiz's Probllama report, a fix committed in about four hours, is the fastest documented in this post.

If you self-host open-weight models, your inference server is likely a bigger line item in your vulnerability backlog than every vendor SDK you use combined. Treat it like any other internet-facing service: authenticate it, bind it to a private interface, and update it on a schedule.

OpenAI vs Anthropic, ChatGPT vs Claude, Gemini vs ChatGPT: how do the pairings compare?

On disclosed CVEs, Anthropic (43) has more than OpenAI (14), and OpenAI has slightly fewer than Google (16). Both gaps come mainly from coding agents: Claude Code has 35 CVEs, Codex 10 and Gemini CLI 2. For the hosted chat products, ChatGPT, Claude and Gemini have each had fixed issues, most without CVEs.

Search interest clusters around a few head-to-head questions, so here are the direct answers from the data.

OpenAI vs Anthropic security. Anthropic has 43 CVEs to OpenAI's 14. Anthropic published 95% of its own, OpenAI 57%. Both are CNAs, both run public bug bounties, both serve a security.txt. The gap is almost entirely Claude Code versus Codex: Claude Code is open about its advisories and has drawn sustained research. Neither vendor's count says which company writes safer code.

ChatGPT vs Claude security. As hosted chat products, both have small CVE footprints. ChatGPT has two CVEs in its apps (the macOS cleartext storage issue and an inline SVG rendering issue) and two documented exposures without CVEs (March 2023 and ShadowLeak). Claude's hosted service has no CVEs; its CVEs are in Claude Code, Claude Desktop and the SDKs. Koi Security's Claude Desktop extension report was fixed without a CVE.

Gemini vs ChatGPT vulnerabilities. Google has 16 CVEs across Gemini and its agent tooling, OpenAI 14. Google's are more often Critical (5 against 2). Both have had prompt injection exfiltration paths fixed without a CVE: the Gemini Trifecta and ShadowLeak.

Copilot vs Gemini security. Microsoft has 75 CVEs to Google's 16, and 15 of Microsoft's are prompt injection in Copilot services. That reflects Microsoft's policy of filing cloud CVEs more than a difference in how either assistant handles untrusted content.

Meta Llama vs the rest. Meta's AI CVE record is small (10) and centred on ExecuTorch and Llama Stack, because Meta ships open weights and infrastructure rather than a hosted API business. The Llama models themselves have no CVEs; the risks of running them sit in whichever inference server you choose.

Does a high count mean an insecure vendor?

No. In this dataset a high CVE count mostly indicates a large installable product surface, a vendor that runs its own CNA, and a policy of disclosing cloud fixes. The lowest counts belong to vendors that ship few installable tools or publish little. Neither end of the table measures how secure a vendor's engineering is.

The concession runs the other way too. A vendor cannot be excused because its count reflects transparency. Claude Code's 35 CVEs and Microsoft 365 Copilot's 29 are real vulnerabilities that users had to be protected from, and the repeated pattern in coding agents (approval bypass through a command the allowlist trusted) suggests the industry's first approach to agent permissions was too permissive. The useful reading is both at once: these vendors found and fixed a lot, and the products they ship are a new and active attack surface.

What a buyer can take from the comparison is narrower and more useful than a ranking. A vendor that publishes advisories, assigns CVEs and states fixed versions is one you can track automatically. A vendor whose CVEs come from third parties, or who has none, needs a different kind of diligence: questions in a security review, contractual notification terms, and monitoring of their status and trust pages.

What should a team building on these products measure instead?

Your risk is not the vendor's total. It is the subset of these CVEs that hits what you run, filtered by whether you can be attacked through it, weighted by how long you stay exposed. In practice that means an inventory of AI components, reachability for each match, and fix latency.

Decision table: the right metric for the question

Question you are askedBad metricBetter metric
"Which AI vendor is most secure?"Raw CVE count per vendorDisclosure maturity (Table 5) plus the CVE profile of the specific products you use
"How exposed are we?"Total CVEs across the vendors we useCVEs matching the AI SDK, agent and runtime versions in our SBOMs
"What do we fix first?"CVSS aloneReachable from our code, known exploited, high EPSS, network exposed
"Are our developers' coding agents safe?"Vendor reputationAgent versions on laptops and CI, update lag, and what each agent may run without approval
"Can we trust this vendor's hosted service?"Absence of CVEsPublic incident history, advisory feed, CNA status, contractual notification terms
"Are we getting better?"Number of open vulnerabilitiesMedian and 90th percentile time from fix published to fix deployed, per AI component

A metric set for AI dependencies

Exposure: distinct vulnerable versions of AI SDKs, agent frameworks, MCP SDKs and inference servers in production SBOMs.

Reachability-adjusted exposure: the same, limited to vulnerabilities whose vulnerable function your code actually reaches.

Agent fleet lag: for each coding agent your developers use, the share of installs more than one release behind.

Model provenance: the share of loaded model files that were scanned before use, by format (pickle, safetensors, GGUF, ONNX).

Connector coverage: the share of MCP servers and connectors in use that sit behind a policy gateway.

Fix latency: time from a fixed version being published to it running everywhere, by component.

Step by step: reproduce the numbers yourself

Everything here runs on macOS or Linux with curl, jq and Python 3. None of it needs an API key.

  1. List the packages that belong to a vendor. Start from the vendor's GitHub organisation and its package registry pages. For Anthropic that means anthropic on PyPI, @anthropic-ai/sdk, @anthropic-ai/claude-code and @anthropic-ai/sandbox-runtime on npm, plus the MCP SDKs if you include the MCP project.

  2. Pull every advisory OSV knows for one package. OSV aggregates the GitHub Advisory Database, PyPA, Go and other feeds, so one call covers most of them.

curl -s -X POST https://api.osv.dev/v1/query \
  -d '{"package":{"name":"@anthropic-ai/claude-code","ecosystem":"npm"}}' \
  | jq -r '.vulns[] | [.id, (.aliases // [] | join(",")), .published[0:10], .database_specific.severity, .summary] | @tsv'
  1. Collapse duplicates. One CVE often appears as a GHSA record and a PYSEC record. Key each advisory on its CVE alias when it has one, and on its GHSA ID when it does not.
curl -s -X POST https://api.osv.dev/v1/query \
  -d '{"package":{"name":"vllm","ecosystem":"PyPI"}}' \
  | jq -r '.vulns[] | ([.id] + (.aliases // [])) | (map(select(startswith("CVE-")))[0] // map(select(startswith("GHSA-")))[0] // .[0])' \
  | sort -u | wc -l
  1. Drop malicious-package records. OSV also carries MAL- entries for compromised or typosquatted releases. They describe a supply chain event, not a coding flaw, so count them separately.

  2. Find the hosted-service CVEs that no package database holds. Cloud and desktop products such as Microsoft 365 Copilot or Copilot Studio never appear in OSV. Search the NVD CVE API 2.0 by keyword (keywordSearch=Copilot, keywordSearch=Claude, and so on), then check each candidate's authoritative record on CVE.org, where the CNA names the affected vendor and product:

curl -s https://cveawg.mitre.org/api/cve/CVE-2025-32711 \
  | jq '{id: .cveMetadata.cveId, cna: .cveMetadata.assignerShortName,
         published: .cveMetadata.datePublished[0:10],
         affected: [.containers.cna.affected[] | {vendor, product}],
         cwe: [.containers.cna.problemTypes[]?.descriptions[]?.cweId]}'

Keep a record only when the affected vendor is the vendor you are counting. Records where MITRE assigned the CVE often say n/a for vendor and product, so read the description for those.

  1. Check who can assign CVEs. The CVE program publishes its partner list as JSON. Each entry includes the CNA's scope, advisory page and disclosure policy.
curl -s https://raw.githubusercontent.com/CVEProject/cve-website/dev/src/assets/data/CNAsList.json \
  | jq -r '.[] | select(.shortName | test("OAI|Anthropic|Google|microsoft|Meta|AMZN")) | [.shortName, .organizationName] | @tsv'
  1. Check security.txt. RFC 9116 puts it at a fixed path.
for d in openai.com anthropic.com google.com www.microsoft.com meta.com mistral.ai x.ai deepseek.com cohere.com aws.amazon.com; do
  printf '%s ' "$d"; curl -s -o /dev/null -w '%{http_code}\n' "https://$d/.well-known/security.txt"
done

A 200 status is not enough on its own. Some single-page apps return their HTML shell for any path, so confirm the body contains a Contact: line.

  1. Measure your own exposure instead of the industry's. Generate an SBOM for each repository and image you ship and match it against the same data. With the Safeguard CLI:
safeguard scan --dir .                 # a repository
safeguard scan --image myapp:latest    # a container image, including OS packages
safeguard scan --sbom sbom.cdx.json    # an existing SBOM

Results are written as JSON under ./output, so you can filter them to the AI packages in this post with the same jq approach.

  1. Keep known vulnerable versions from coming back. A minimal GitHub Actions gate with the free Gold scan action:
name: dependency-gate
on: [pull_request]
jobs:
  gold-scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      # pin to a released tag of the action
      - uses: safeguard-sh/gold-scan-action@main

Check the action's README for the input that decides whether the build fails on known exploited vulnerabilities only, on any known vulnerability, or never, and pin the version you adopt.

How we measured

Scope. Ten LLM vendors: OpenAI, Anthropic, Google, Microsoft, Meta, Mistral AI, xAI, DeepSeek, Cohere and Amazon. For each, we counted vulnerabilities in the vendor's own generative AI products: SDKs, CLIs and coding agents, desktop, mobile and browser apps, IDE plugins and extensions, agent frameworks, MCP servers the vendor publishes, hosted AI services, and the ML platforms each cloud vendor uses to serve models (Azure Machine Learning and Azure AI Foundry, SageMaker and Bedrock, Vertex AI). Products built by third parties on a vendor's API were excluded. The MCP project and three open-model runtimes (vLLM, llama.cpp, Ollama) are reported as separate rows. Nine Azure Machine Learning CVEs and eleven SageMaker CVEs are included; we excluded non-generative Azure AI services such as Face, Language and AI Search.

Package advisories. We queried the OSV API on 16 November 2026 for 88 packages across PyPI, npm, Go, crates.io, Maven and NuGet published by the vendors, the MCP project and the runtimes, from openai and @openai/codex to anthropic, @anthropic-ai/claude-code, @google/gemini-cli, google-adk, semantic-kernel, llama-stack, executorch, mistralai, cohere, strands-agents and bedrock-agentcore. OSV aggregates the GitHub Advisory Database, PyPA and Go feeds. We collapsed duplicates on the CVE alias (or the GHSA ID when no CVE exists) and set aside malicious-package records as a separate category.

Hosted and desktop products. Cloud services, desktop apps and IDE extensions are not in package databases. We ran 42 keyword searches against the NVD CVE API 2.0 on 16 November 2026 (for example Copilot, Claude, Gemini, Codex, ChatGPT, ExecuTorch, Bedrock, Kiro, Grok, DeepSeek, Mistral, Cohere, Ollama, vLLM, llama.cpp, Model Context Protocol), which returned 2,091 candidate CVEs. We then fetched every candidate's authoritative record from the CVE.org API and kept it only when the CNA-supplied affected vendor and product matched the vendor's own product. For 151 records where MITRE or another CNA left vendor and product as "n/a", we read the description and classified by hand; 26 were added and the rest were third-party products that only mention a vendor's name (for example OpenAirInterface, a 5G project unrelated to OpenAI).

Attribution choices. GitHub Copilot counts toward Microsoft. MCP SDKs and reference servers are shown as their own row because Anthropic's published CNA scope covers them while the project itself is now community governed. DeepSeek Harness counts toward DeepSeek because it is published under the deepseek-ai GitHub organisation.

Severity. CNA CVSS base score from the CVE.org record, preferring v3.1, then v4.0, then the ADP score on the same record. Bands: Critical 9.0 and above, High 7.0 to 8.9, Medium 4.0 to 6.9, Low below 4.0. CVSS v3.1 and v4.0 scores are not directly comparable, and CNAs score differently, so bands are approximate.

Weakness families. CWE IDs from the CNA record (or the ADP container when the CNA gave none), grouped into ten families. Prompt injection is assigned only when the record uses CWE-1427, when the CNA calls it "AI command injection", or when a hosted assistant's command injection leads to information disclosure; a CVE that merely mentions prompt injection as the delivery route keeps its underlying CWE family. CVEs without a CWE fall into "other", which is why CVE-2024-50050 (Llama Stack, pickle) appears there.

Layer and product line. Assigned from the affected product named on the record, using a fixed set of patterns per vendor, then checked by hand.

Normalisation. Three denominators: the 2026 calendar year to date, the number of product lines with at least one CVE, and the share of CVEs published through the vendor's own channel (its own CNA, or GitHub's CNA acting on an advisory the vendor published on its own repository).

Disclosure maturity. Checked on 16 November 2026: /.well-known/security.txt on each vendor's main domains (requiring a real Contact: line, since some single-page apps return HTML for any path), the CVE program's published CNA list, vendor policy and bug bounty pages, and published GitHub repository advisories.

Notable cases and time to fix. Each case was verified against its CVE.org record and a primary source (the vendor's advisory or the researcher's write-up) where one was retrievable; where a primary page blocked automated access we used two independent secondary reports and say so. Time to fix is shown only where both a report date and a fix date were published.

Not done. No weighting by exploitability or deployment numbers, no adjustment for research attention, and no attempt to count silent fixes. CVE publication lags fixes, so recent months will rise as records are filed. The NVD keyword search depends on NVD's index; a CVE.org record that NVD had not yet processed on the query date would be missed unless OSV carried it.

Where sources disagree. Severity for CVE-2024-50050 (CNA 6.3, independent 9.3). The number of indexed Grok conversations (Forbes reported more than 370,000; another report cited about 4,500). The Llama Stack report date (24 or 29 September 2024, depending on the source). The EchoLeak report date (January 2025 per secondary reporting; no primary page we could retrieve states it).

Common mistakes when comparing AI vendors on security

Reading a raw CVE count as a safety ranking. A vendor with its own CNA, a public advisory feed and an open-source CLI will always out-count a vendor that runs everything server-side and publishes nothing. The count measures disclosure as much as it measures defects.

Counting only CVEs. Several of the most consequential events in this dataset (the DeepSeek database, the ChatGPT Redis exposure, the Grok shared links, the xAI key on GitHub) have no CVE at all. A CVE-only view misses every hosted-service data exposure.

Treating the MCP project as nobody's. Its SDKs and reference servers sit inside Anthropic's CNA scope, and many teams install them alongside vendor SDKs from every company in this post. Track them as their own row in your inventory.

Assuming a server-side fix needs no action. EchoLeak was fixed by Microsoft with nothing for customers to install. The Amazon Q extension compromise, the Ollama path traversal and every Claude Code CVE required users to update a local install. Know which kind you are looking at.

Prioritising by count instead of reachability. Most advisories in an AI dependency tree hit code paths you never call. A deserialization bug in a distributed-inference transport does not reach an application that only uses the tokenizer.

Fix it without waiting for upstream

The vendors in this comparison fixed nearly every case above. The fix landing in a vendor's repository is not the same as the fix running in your build. Your exposure is the gap between those two moments, plus the longer gap when no fix exists yet.

Teams building on these SDKs, CLIs and runtimes inherit their vulnerabilities in three ways. A vulnerable package sits in a lockfile. A vulnerable tool sits on a developer laptop or a CI runner. A vulnerable runtime or model file sits in a container. Safeguard covers all three from one platform, so the team that reads this comparison can act on it without stitching together five tools.

Flow diagram: AI SDKs, inference servers, model files, packages and coding agents on the left; Safeguard SCA with reachability, image scanning, AI model scanning, Package Firewall, Guard, AI Gateway and MCP Guardrail in the middle; fix pull requests, Gold patched builds and blocks on the right

Figure 8. How Safeguard covers the AI stack end to end. Safeguard capabilities as offered on 16 November 2026. Safeguard did not discover the vulnerabilities discussed in this post.

1. Know what you run: SCA with reachability across AI SDK dependencies

Safeguard's software composition analysis builds a full transitive inventory of every repository and image, including the AI layer: openai, anthropic, @anthropic-ai/claude-code, @google/gemini-cli, google-adk, semantic-kernel, llama-stack, vllm, the MCP SDKs, and the long tail of agent frameworks pulled in beside them. SBOMs export as SPDX or CycloneDX. Reachability analysis then asks the question raw counts cannot: does your code actually call the vulnerable path? A pickle deserialization flaw in a vLLM distributed-inference transport matters a great deal if you run multi-node serving and very little if you only import the tokenizer. Reachability cuts the list to the vulnerabilities you can be attacked through, typically removing 60 to 80 percent of the noise.

2. Scan the models, not just the code

Several of the most severe cases in this dataset share one root cause: a model file or a serialized object is treated as data when it can execute code. Safeguard's AI model scanning checks model artifacts in pickle, safetensors, GGUF and ONNX formats for malicious or unsafe content before anyone loads them, and AI-BOM discovery inventories the models, SDKs and agents across your estate. That inventory answers the first question in any model supply chain review: which models do we actually run?

3. Guard the agent at runtime

Coding agents and agent frameworks account for most of the 2025 and 2026 growth in this data. Many of those vulnerabilities share a pattern: a prompt or repository file convinces an agent to run a command, read a file or call a URL it should not. Guard, Safeguard's runtime protection layer, inspects system calls, library calls and network egress, and blocks PII and secret exfiltration on the egress path. The AI Gateway in front of model calls adds guardrails against prompt injection, jailbreaks and PII leaving in prompts or responses. A sed or find command smuggled past an approval prompt still has to get its data off the machine, and that egress path is where Guard sits.

4. Put a policy in front of every MCP server and connector

The MCP rows in this comparison are a reminder that the connector layer is now its own attack surface: DNS rebinding, missing authentication between a client and proxy, path validation bypasses, OAuth token confusion. Safeguard's MCP Guardrail gives each MCP server a proxy URL with policy enforcement in front of it, adding under 10 ms of latency, and the MCP security SDK brings the same checks into servers you build yourself. Agent identity tracking records which agent holds which access, at what level, so an over-scoped token is visible before it is abused.

5. Fix pull requests and Gold patched packages

When a fixed version exists, Safeguard's Griffin models open a tested, policy-gated pull request that moves the dependency to the safe version, on one repository or in bulk across every project carrying it. Autonomous remediation ships switched off, with every category set to Manual, so your team decides what runs on its own.

When upstream has not shipped a fix, or will not, Gold closes the gap. gold.safeguard.sh is a free, no-login directory where you can look up any CVE in this post and see the affected and fixed versions with CVSS, EPSS and KEV status. Gold Registry at registry.safeguard.sh serves hardened, drop-in builds held to zero exploitable critical or high CVEs and zero malware, signed with Sigstore and carrying SLSA provenance. Where upstream is slow, Safeguard forks the package, patches the vulnerability and publishes a hardened Gold build, so a maintainer's release schedule stops being your exposure window.

6. Stop malicious releases at install time

The May 2026 malicious mistralai releases show that a trusted vendor name on a package is not proof the package is safe. Safeguard's Package Firewall sits between developers and the public registries for npm and pip and blocks known malicious packages, typosquats and dependency confusion attempts before they install.

Your situationWhat to use
You need to know which AI SDKs, agents and runtimes you run, and whereSCA inventory and SBOM export across repositories and images
A CVE in this post affects a package you useReachability to confirm it matters, then a Griffin fix pull request
Upstream has no fix yetA hardened Gold build from Gold Registry
Your team downloads models from public hubsAI model scanning for pickle, safetensors, GGUF and ONNX
Developers run coding agents with shell accessGuard runtime protection on execution and egress, plus AI Gateway guardrails
You connect agents to internal systems through MCPMCP Guardrail proxy and agent identity tracking
You worry about a compromised vendor releasePackage Firewall on npm and pip installs

One honest limit: Safeguard cannot patch a vendor's hosted service. When the vulnerable code runs inside Microsoft 365, ChatGPT or Gemini in the cloud, the vendor's server-side fix is the only fix. What you control is everything you install, build, connect and run around those services, and that is where most of the vulnerabilities in this dataset live.

Try it yourself

Click through these interactive Safeguard demos to see the steps described in this guide.

Test yourself

Which vendor has the most CVEs in its AI products, and what mostly explains it?

Microsoft has 75, ahead of Amazon (53) and Anthropic (43), as of 16 November 2026. Much of the gap comes from its policy since June 2024 of issuing CVEs for critical cloud-service vulnerabilities even when customers have nothing to install; 55 of its 75 are in hosted services and apps.

Does a high CVE count mean an insecure vendor?

No. In this data a high count mostly reflects a large installable product surface, a vendor that runs its own CVE Numbering Authority, and a policy of disclosing cloud fixes. A count of zero, like xAI's, only means nothing has been filed.

Why does the post show the MCP project as a separate row?

Its SDKs and reference servers sit inside Anthropic's CNA scope and add 55 CVEs. They are shown separately rather than folded into Anthropic's count or left out.

How fast is the count growing?

Very fast and almost all recently: 185 of the 241 vendor CVEs (77 percent) were published in 2026, and 100 of them (41 percent) are in coding agents, CLIs and IDE extensions.

What does disclosure channel tell you about a vendor?

It shows who told you. Google, Microsoft, Amazon, Anthropic and Meta published nearly all of their own CVEs themselves, while every CVE for Mistral AI, DeepSeek and Cohere was filed by a third party. That matters to a team that wants one advisory feed per supplier.

How quickly do vendors fix reported vulnerabilities?

Documented fixes took from three days (Ollama's Probllama) to roughly four months (EchoLeak), with most between two and seven weeks. There are too few cases to rank vendors on speed.

What should a team measure instead of raw CVE counts?

The subset of CVEs that hits what you run, filtered by reachability and weighted by how long you stay exposed. In practice that means an inventory of AI components, reachability for each match, and fix latency.

FAQ

Which AI company has the most security vulnerabilities?

By count of CVEs in its own AI products, Microsoft has the most: 75 on 16 November 2026, followed by Amazon (53) and Anthropic (43). The count largely reflects how much installable software each vendor ships and whether it assigns CVEs to cloud-service fixes. Microsoft does both. Open-model runtimes carry more than any vendor, with vLLM at 140.

Is OpenAI or Anthropic more secure?

The data does not say either is more secure. Anthropic has 43 CVEs and OpenAI 14, but 35 of Anthropic's are in Claude Code, an open-source coding agent with a public advisory feed, and 10 of OpenAI's are in Codex. Both are CVE Numbering Authorities, both run public bug bounties and both publish advisories. Anthropic published 95% of its CVEs itself, OpenAI 57%.

Is ChatGPT or Claude safer to use?

As hosted chat services, both have small CVE records. ChatGPT has two CVEs in its apps and two notable fixed exposures without CVEs: the March 2023 Redis bug and ShadowLeak in 2025. Claude's hosted service has no CVEs; Anthropic's CVEs are in Claude Code, Claude Desktop and its SDKs. For either, the bigger risk usually sits in the tools and connectors you add around the chat.

How many CVEs does Claude Code have?

Claude Code has 35 CVEs as of 16 November 2026, the most of any single vendor product in this comparison. Most are High-severity approval or sandbox bypasses, where a command the read-only allowlist trusted, a symlink, a git worktree or a repository settings file could run or write before the user approved. Each was fixed in a subsequent Claude Code release, and Anthropic published nearly all of them as GitHub advisories.

What was EchoLeak (CVE-2025-32711)?

EchoLeak was a zero-click indirect prompt injection in Microsoft 365 Copilot, reported by Aim Labs. A crafted email could make Copilot pull sensitive data from the user's context and send it to an attacker without any click. Microsoft scored it 9.3, fixed it server-side, published the CVE on 11 June 2025 and said customers did not need to take any action.

Did DeepSeek have a data breach?

DeepSeek had a significant data exposure. On 29 January 2025, Wiz Research reported a publicly accessible DeepSeek ClickHouse database that exposed more than a million log lines, including chat history, secret keys and backend details. DeepSeek secured it promptly after the report. There is no CVE for it. DeepSeek's 15 CVEs, all filed by third parties, are in DeepSeek Harness and its models and chat service.

Why does xAI have no CVEs?

xAI ships few installable products, has no CVE Numbering Authority and publishes no security advisories, so there is little for a CVE to attach to and no in-house route to request one. That does not mean nothing has happened. An employee's API key sat exposed on GitHub for about two months in 2025, and Forbes reported that over 370,000 shared Grok conversations were indexed by search engines.

Are Ollama and vLLM safe to self-host?

They can be run safely, but they need the care of any network service. vLLM has 140 CVEs and Ollama 34, many of them denial of service, memory safety in model parsing, or missing authentication on the API. Bind them to a private interface, put authentication in front, scan model files before loading them, and update on a schedule. Ollama's API had no authentication through 0.12.3.

Quick check

Microsoft has the most CVEs among LLM vendors. What explains most of the gap? Its policy, since June 2024, of publishing CVEs for critical cloud-service vulnerabilities even when customers have nothing to install. 55 of its 75 CVEs are in hosted services and apps such as Microsoft 365 Copilot.

Which disclosed vulnerability class appears in almost every vendor's coding agent? Approval and sandbox bypass: content in a repository, such as a settings file, Git configuration, MCP configuration or a crafted command, running before or around the user's approval. It appears in Claude Code, Codex CLI, Gemini CLI, GitHub Copilot, Kiro and Mistral Vibe.

Sources

CVE.org CVE Services API, records fetched 16 November 2026: https://cveawg.mitre.org/api/cve/CVE-2025-32711

CVE Program partner (CNA) list, JSON published by the CVE program: https://github.com/CVEProject/cve-website/blob/dev/src/assets/data/CNAsList.json

CVE Program, List of Partners: https://www.cve.org/PartnerInformation/ListofPartners

OSV API documentation: https://google.github.io/osv.dev/api/

GitHub Advisory Database: https://github.com/advisories

CVE-2025-32711 (EchoLeak), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-32711

CVE-2024-38206 (Copilot Studio SSRF), CVE.org: https://www.cve.org/CVERecord?id=CVE-2024-38206

CVE-2025-53773 (GitHub Copilot and Visual Studio), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-53773

CVE-2026-25592 (Semantic Kernel), CVE.org: https://www.cve.org/CVERecord?id=CVE-2026-25592

CVE-2024-50050 (Llama Stack), CVE.org: https://www.cve.org/CVERecord?id=CVE-2024-50050

CVE-2025-30405 (ExecuTorch), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-30405

CVE-2025-52882 (Claude Code IDE extensions), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-52882

CVE-2025-54794 (Claude Code), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-54794

CVE-2025-54795 (Claude Code), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-54795

CVE-2026-96620 (Claude Agent SDK for Python), CVE.org: https://www.cve.org/CVERecord?id=CVE-2026-96620

CVE-2025-49596 (MCP Inspector), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-49596

CVE-2024-40594 (ChatGPT for macOS), CVE.org: https://www.cve.org/CVERecord?id=CVE-2024-40594

CVE-2025-59532 (Codex CLI), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-59532

CVE-2025-61260 (Codex CLI), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-61260

CVE-2026-19593 (Codex Desktop), CVE.org: https://www.cve.org/CVERecord?id=CVE-2026-19593

CVE-2026-12537 (Gemini CLI), CVE.org: https://www.cve.org/CVERecord?id=CVE-2026-12537

CVE-2026-4810 (Google Agent Development Kit), CVE.org: https://www.cve.org/CVERecord?id=CVE-2026-4810

CVE-2025-8217 (Amazon Q Developer for VS Code), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-8217

CVE-2026-116843 (Mistral Vibe), CVE.org: https://www.cve.org/CVERecord?id=CVE-2026-116843

CVE-2024-37032 (Ollama, Probllama), CVE.org: https://www.cve.org/CVERecord?id=CVE-2024-37032

CVE-2025-63389 (Ollama missing authentication), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-63389

CVE-2025-47277 (vLLM PyNcclPipe), CVE.org: https://www.cve.org/CVERecord?id=CVE-2025-47277

CVE-2023-28858 (redis-py), CVE.org: https://www.cve.org/CVERecord?id=CVE-2023-28858

MSRC, Toward greater transparency: Unveiling Cloud Service CVEs (27 June 2024): https://msrc.microsoft.com/blog/2024/06/toward-greater-transparency-unveiling-cloud-service-cves

OpenAI, March 20 ChatGPT outage: Here's what happened: https://openai.com/index/march-20-chatgpt-outage/

OpenAI, Mixpanel security incident notice (November 2025): https://openai.com/index/mixpanel-incident/

Help Net Security, A bug revealed ChatGPT users' chat history, personal and billing data (27 March 2023): https://www.helpnetsecurity.com/2023/03/27/chatgpt-data-leak/

SiliconANGLE, Mixpanel breach compromises OpenAI API users' account data (27 November 2025): https://siliconangle.com/2025/11/27/mixpanel-breach-compromises-openai-api-users-account-data/

9to5Mac, ChatGPT for macOS conversations stored in plain text (3 July 2024): https://9to5mac.com/2024/07/03/chatgpt-macos-conversations-plain-text

Radware, ShadowLeak: https://www.radware.com/blog/posts/shadowleak/

The Hacker News, ShadowLeak zero-click flaw leaks Gmail data (September 2025): https://thehackernews.com/2025/09/shadowleak-zero-click-flaw-leaks-gmail.html

Wiz Research, Wiz Research Uncovers Exposed DeepSeek Database Leaking Sensitive Information (29 January 2025): https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak

Wiz Research, Probllama: Ollama Remote Code Execution Vulnerability CVE-2024-37032 (24 June 2024): https://www.wiz.io/blog/probllama-ollama-vulnerability-cve-2024-37032

Oligo Security, CVE-2024-50050: Critical Vulnerability in meta-llama/llama-stack: https://oligo.security/blog/cve-2024-50050-critical-vulnerability-in-meta-llama-llama-stack

Oligo Security, More Models, More ProbLLMs (Ollama): https://www.oligo.security/blog/more-models-more-probllms

The Hacker News, Meta's Llama Framework Flaw Exposes AI Systems to Remote Code Execution Risks (January 2025): https://thehackernews.com/2025/01/metas-llama-framework-flaw-exposes-ai.html

Aim Labs, EchoLeak: https://www.aim.security/lp/aim-labs-echoleak-blogpost

The Hacker News, Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data (June 2025): https://thehackernews.com/2025/06/zero-click-ai-vulnerability-exposes.html

Tenable, SSRFing the Web with the Help of Copilot Studio (20 August 2024): https://www.tenable.com/blog/ssrfing-the-web-with-the-help-of-copilot-studio

Tenable, The Trifecta: How Three New Gemini Vulnerabilities in Cloud Assist, Search Model, and Browsing Allowed Private Data Exfiltration (30 September 2025): https://www.tenable.com/blog/the-trifecta-how-three-new-gemini-vulnerabilities-in-cloud-assist-search-model-and-browsing

Tracebit, Code Execution Through Deception: Gemini AI CLI Hijack: https://tracebit.com/blog/code-exec-deception-gemini-ai-cli-hijack

AWS Security Bulletin AWS-2025-015, Amazon Q Developer Extension for Visual Studio Code: https://aws.amazon.com/security/security-bulletins/AWS-2025-015/

Cymulate, InversePrompt: Turning Claude Against Itself (CVE-2025-54794 and CVE-2025-54795): https://cymulate.com/blog/cve-2025-547954-54795-claude-inverseprompt/

Koi Security, PromptJacking: the critical RCE in Claude Desktop: https://www.koi.ai/blog/promptjacking-the-critical-rce-in-claude-desktop-that-turn-questions-into-exploits

Infosecurity Magazine, Claude Desktop Extensions Vulnerable to Web-Based Prompt Injection: https://www.infosecurity-magazine.com/news/claude-desktop-extensions-prompt/

KrebsOnSecurity, xAI Dev Leaks API Key for Private SpaceX, Tesla LLMs (1 May 2025): https://krebsonsecurity.com/2025/05/xai-dev-leaks-api-key-for-private-spacex-tesla-llms/

Forbes (Australia edition), xAI published hundreds of thousands of Grok chatbot conversations (August 2025): https://www.forbes.com.au/news/innovation/xai-published-hundreds-of-thousands-of-grok-chatbot-conversations/

TechCrunch, Meta fixes bug that could leak users' AI prompts and generated content (15 July 2025): https://techcrunch.com/2025/07/15/meta-fixes-bug-that-could-leak-users-ai-prompts-and-generated-content

GitHub advisory GHSA-wx9m-wx4f-4cmg, malicious dropper in mistralai 2.4.6: https://github.com/advisories/GHSA-wx9m-wx4f-4cmg

GitHub advisory GHSA-jgg6-4rpr-wfh7, compromised @mistralai npm packages: https://github.com/advisories/GHSA-jgg6-4rpr-wfh7

The Register, Google declares AI bug hunting season open (7 October 2025): https://www.theregister.com/2025/10/07/google_ai_bug_bounty/

OpenAI security.txt: https://openai.com/.well-known/security.txt

OpenAI CVE assignment policy: https://openai.com/policies/openai-cve-assignment-policy/

OpenAI bug bounty on Bugcrowd: https://bugcrowd.com/openai

Anthropic responsible disclosure policy: https://www.anthropic.com/responsible-disclosure-policy

Anthropic bug bounty on HackerOne: https://hackerone.com/anthropic

Google security.txt: https://google.com/.well-known/security.txt

Microsoft security.txt: https://www.microsoft.com/.well-known/security.txt

Meta bug bounty: https://bugbounty.meta.com/

Mistral AI security.txt: https://mistral.ai/.well-known/security.txt

xAI security.txt: https://x.ai/.well-known/security.txt

Cohere Trust Center: https://trustcenter.cohere.com/

Amazon Vulnerability Research Program on HackerOne: https://hackerone.com/amazonvrp

AWS vulnerability reporting: https://aws.amazon.com/security/vulnerability-reporting/

Ollama SECURITY.md: https://github.com/ollama/ollama/blob/main/SECURITY.md

vLLM SECURITY.md: https://github.com/vllm-project/vllm/blob/main/SECURITY.md

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

We email you a link to confirm, and only then add you to the list. Unsubscribe in one click. Privacy

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.