Samuel Arendt

LiteLLM: CVE-2026-40217 in the guardrail test endpoint added to CISA's exploited vulnerabilities catalog

On November 18, 2026, CISA added CVE-2026-40217, a LiteLLM flaw, to its Known Exploited Vulnerabilities (KEV) catalog. LiteLLM is the open source proxy many teams use as a single entry point for calling models from several providers, managing keys and tracking what agents and internal tools spend. A KEV listing means the agency has evidence of real exploitation. US federal agencies have until December 2 to fix it.

The flaw in the guardrail test endpoint

The project's advisory, published on May 7, describes the problem in the POST /guardrails/test_custom_code endpoint. It runs user-supplied Python inside a sandbox the project wrote itself, and bytecode manipulation techniques let an attacker escape it and run arbitrary code in the proxy process. In the default Docker image, that process runs as root.

Affected versions run from 1.81.8 up to, but not including, 1.83.10. The advisory's patched-versions field says 1.83.10, while its own text says the fix landed in 1.83.11; anyone still in that range should go straight to a version newer than both.

The advisory rates the flaw as moderate (CVSS 4.0 score of 5.9) because exploitation requires a proxy admin credential. CISA does not say how attackers are getting that access.

Who needs to look at this

Teams that host their own LiteLLM proxy, in a container or on a server, on an April release or older. That includes setups built to centralize access for coding agents and product AI features that were never updated afterwards. The advisory for another project flaw (CVE-2026-49468) says LiteLLM Cloud customers were not affected by it; the CVE-2026-40217 advisory does not say either way.

The admin credential requirement is less of a barrier than it looks. The project published other flaws this year that lead to that level of access. CVE-2026-49468, fixed in 1.84.0, allowed unauthenticated access to management routes through a forged Host header when nothing in front of the proxy validated that header. CVE-2026-114612, from September and scored 9.9, let an ordinary internal user pass as an admin. An install frozen in April is still exposed to the first; anything between 1.87.0 and the versions fixed in September (such as 1.100.4) may be in range of the second.

A proxy like this holds the API keys for every model provider the company uses. Code execution in that process gives access to those keys and to whatever passes through the requests.

What to do now

  1. Find out whether LiteLLM is running anywhere in the company and on which version. Proxies set up by one team for a pilot often stay out of the inventory.
  2. Upgrade to the current stable release, which also covers the privilege flaws above. Outside US federal agencies, the December 2 deadline is a useful reference; with confirmed exploitation, I would not wait for it.
  3. If the upgrade will take time, block POST /guardrails/test_custom_code at the reverse proxy or API gateway, the workaround given in the advisory.
  4. Rotate the proxy master key and the provider keys it stores if the admin panel was exposed to the internet on an affected version, and review the logs for calls to that endpoint.

This kind of service is usually installed by a project and forgotten after delivery; the question of who looks after the system once the project ends applies to AI infrastructure too.

What CISA and the project have not said yet

CISA lists use in ransomware campaigns as unknown and does not say who is exploiting the flaw or since when. It is also unclear whether the attacks chain CVE-2026-40217 with one of the privilege flaws or start from leaked admin credentials. The project has not published indicators of compromise so far.

If your company is weighing something similar, reach me on LinkedIn or WhatsApp.

References

Clear sky seen between the dark vertical bars of a window, with one horizontal crossbar.

Photo by Samuel Dedman on Unsplash

#AI Agents, #Operations, #Radar