Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
An integer overflow in a vLLM CUDA kernel could copy one user's response into another's in the same batch. It led eight vLLM advisories published in September 2026.
GHSA-gqvg-gmmx-x4hm: with MLFLOW_ALLOW_PICKLE_DESERIALIZATION=False, a model tagged as statsmodels still unpickled and ran code. Fixed in MLflow 3.15.0.
Nine Docling advisories published 5 to 7 October 2026 cover local file reads, SSRF, plugin code running despite a block, and tiny files that exhaust memory. Fixed by 2.132.0.
OpenCode's Console breach sent teams hunting for provider keys. Use the OpenAI, Anthropic and Google admin APIs to list every key, tag an owner and retire the unused ones.
The tensorlake worm committed .claude/settings.json and .vscode/tasks.json into victims' repos. Find files like these across your org, then lock down who can change them.
CISA added Langflow's exec_globals RCE, CVE-2026-0770, to KEV on 21 July 2026. The ZDI advisory lists no vendor fix, so reduce exposure first.
Langflow before 1.9.1 let any logged in user run flows owned by others. CISA put CVE-2026-55255 on KEV on 7 July 2026 with a three day deadline.
Cato AI Labs disclosed DuneSlide on 1 July 2026. Two Cursor bugs let injected text break out of the terminal sandbox. Cursor 3.0 fixes both.
Claude Code auto-approved every path on huggingface.co in WebFetch. Attackers could use download counts as a covert channel. Fixed in 2.1.163.
Weekly insights on software supply chain security, delivered to your inbox.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.
