High 5.9Being exploited
BerriAI · LiteLLM (AI gateway and proxy) · CVE-2026-40217
LiteLLM AI gateway flaw used by attackers to turn limited keys into full admin access
LiteLLM is an open-source gateway that many teams put between their applications and AI services such as OpenAI and Anthropic, and it holds the keys to those services. A flaw in a feature for testing custom guardrail code lets someone escape its sandbox and run their own code in the gateway, as root in the default Docker image. The maintainers rate it moderate (CVSS 5.9) because in default setups it needs an admin credential, but Google's threat intelligence team says it saw an attacker use it to turn limited keys into full administrative credentials. CISA added it to its exploited catalogue on 18 November.
- Affected
- The litellm Python package and proxy from version 1.81.8 up to, but not including, 1.83.10, according to the maintainers' advisory. Gateways reachable from the internet, or with keys shared widely, are most at risk.
- What to do
- Upgrade LiteLLM to 1.83.11 or later, the version in which the advisory says the sandbox was replaced (its version table lists 1.83.10 as the first patched release, so go to 1.83.11 to be sure), and rebuild any container images. If you cannot upgrade straight away, the maintainers' workaround is to block POST requests to /guardrails/test_custom_code at your reverse proxy or API gateway. If your gateway was exposed on an affected version, rotate its admin key and the AI provider keys it stores, and check provider bills for usage you do not recognise.General information summarised from the vendor's advisory at the time of writing, not professional advice. Check the vendor's current advisory before acting.
- Why it’s here
- Confirmed exploited and newly on CISA's catalogue, in a popular tool for adding AI features to websites and apps, where stolen keys mean someone else's usage on your bill.
