LiteLLM guardrail sandbox escape (CVE-2026-40217)
A sandbox escape in LiteLLM's custom-code guardrail test endpoint that gives code execution in the LLM gateway; disclosed May 2026 and added to CISA's Known Exploited Vulnerabilities catalog on Nov 18 2026.
What happened
LiteLLM’s proxy exposes POST /guardrails/test_custom_code, which runs administrator-supplied Python to test custom guardrails. The GitHub advisory published May 7 2026 says that code ran in a hand-rolled sandbox that could be escaped with bytecode-level techniques, giving arbitrary code execution in the proxy process, which runs as root in the default Docker image. In default configurations the endpoint needs a proxy-admin credential, one reason the advisory rates it only Moderate. The fix replaced the sandbox with RestrictedPython; the advisory’s version table says 1.83.10 fixes it, while its text names 1.83.11. Blocking the endpoint at a reverse proxy is the stated workaround.
On Nov 18 2026 CISA added CVE-2026-40217 to its Known Exploited Vulnerabilities catalog, meaning it has evidence of exploitation in the wild, and gave federal agencies until Dec 2 to patch. CISA did not say who exploited it or how.
Why it matters
LiteLLM sits in front of model providers and holds their API keys. Exposed instances are already targets: the PoeLLM cryptominer infected mostly internet-facing LiteLLM and Ollama servers, and ARTEX agents exploited other AI tools to steal provider keys. Like GPUStack earlier in November, it is an AI-serving tool in the KEV catalog; no source we found says AI was used to find or exploit it.
Timeline
Every dated event in the wiki that involves LiteLLM guardrail sandbox escape (CVE-2026-40217), newest first.
- CISA lists a LiteLLM guardrail sandbox escape (CVE-2026-40217) as exploited (LiteLLM guardrail sandbox escape (CVE-2026-40217))
CISA added the flaw in the open-source LLM gateway to its Known Exploited Vulnerabilities catalog on Nov 18 with a Dec 2 deadline. Per the May 2026 GitHub advisory, the custom-code guardrail test endpoint ran user-supplied Python in a hand-rolled sandbox that could be escaped for code execution in the proxy process, which runs as root in the default Docker image; it needs a proxy-admin credential by default. Versions 1.81.8 to before 1.83.10 are affected. CISA did not say who exploited it, and no source we found says AI was used.
Connections
LiteLLM guardrail sandbox escape (CVE-2026-40217) and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
Concept AI and CVEs: latent flaws, now exploitable
Product LiteLLM