AI Security
In-depth guides and analysis on ai security from the Safeguard engineering team.
100 articles
What Anthropic's S-1 tells security teams about AI vendor risk
Anthropic's S-1 discloses weight theft risk, agent attack surfaces, cloud concentration and export curbs. How to use it in an AI vendor review.
Human Brain vs LLM Context Window: Memory, Attention and AI Security
Working memory holds about four chunks; a context window can hold a million tokens. Where the analogy breaks, and why anything in context can steer a model.
What Is LLMjacking, and How Do We Stop AI API Key Theft?
LLMjacking is running AI models on your account with a stolen key or open endpoint. Inventory keys, scope them, make them short-lived and watch usage.
Can AI agents now hack websites on their own? What defenders change
Yes, within limits. AI agents now run web intrusions, mostly through known, unpatched flaws. Here is what is confirmed and what defenders should change.
Can AI agents escape their sandbox? How to contain the agents you run
Yes. In 2026 lab agents left sandboxes through vulnerable services and open egress. Here is how to contain agents with VMs, egress control and monitoring.
Google Agent Mesh CVE-2026-102503 and CVE-2026-103091: Why Upgrading Isn't Enough
Two critical Agent Mesh (SAM) CVEs, fixed in 0.1.0-alpha.3 and alpha.9. How they work, why a stored wildcard policy survives the upgrade, and a checker.
Envoy AI Gateway MCP Session Hijacking (GHSA-953h-535j-jg2v): The v1.2.0 Fix and the Seed That Can Outlive It
Envoy AI Gateway, now Agent Router, let any caller resume another user's MCP session. v1.2.0 fixes it; a fallback seed can keep the old hole open. Checker.
MCP Guardrail for AI agents: setup, policies and daily use
Set up Safeguard MCP Guardrail: limit which tools agents can see, check every MCP call with allow, deny or monitor rules, and review the audit trail.
How to Choose an AI Model for Security and Compliance Work
Choose by data rules first: training, retention and who can review your data. Then match the job; some security questions need no language model at all.
AI coding agents in regulated industries: residency, audit and approvals
Regulated teams can use AI coding agents if they control where code goes, who approves each action, what gets recorded and which models are allowed.
AI coding agent security: the main risks and the controls that work
AI coding agents add dependencies, run commands, send code to models and read untrusted text. Approvals, scans, policies and MCP limits keep that in check.
AI Coding Agent Approval Workflows: What, Where and Who Approves
Approve AI coding agents at three points: risky actions in the session, the diff in review by someone other than the requester, and the merge gate in CI.
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.