LiteLLM flaws let attackers take over AI gateways and steal API keys

LiteLLM is one of the most widely used open-source AI gateways: a layer that sits between your apps and model providers so teams can switch models, track spend and manage keys in one place. Security researchers have now disclosed flaws that let low-privileged users take over a LiteLLM instance's admin controls and steal the provider API keys stored inside it.

That combination is what makes this serious. A gateway holds the keys to every model account a company uses. Anyone who takes it over can run up bills on those accounts, read traffic that passes through, or reuse the keys elsewhere. One report estimates that 9.6% of internet-facing LiteLLM instances were exposed.

The fix path is standard but urgent: update to a patched release, check who has access to the admin interface, and treat every key that passed through an exposed instance as compromised. The episode is also a reminder that AI infrastructure is now ordinary infrastructure, with ordinary vulnerabilities, and attackers know where the keys are.

Our take

AI gateways are becoming the new secrets vault, and most teams installed theirs as a convenience tool. They deserve the same scrutiny as any system that holds production credentials.