Latest / Decrypted: The UK Cyber Briefing / The AI gateway bug that turned a failed login into a free pass
Transcript
- 0:03This is Decrypted. Monday, the seventh of September.
- 0:06The AI gateway bug that turned a failed login into a free pass.
- 0:11A widely used open-source gateway for running large language model applications has a serious problem.
- 0:18It's called LiteLLM, and for months it had an authentication check that failed the wrong way.
- 0:24When it couldn't verify who was calling, it let them in anyway.
- 0:28Sisa added the flaw, C V E twenty twenty-six, five nine eight two two, to its Known Exploited Vulnerabilitie
- 0:37s catalogue on 2 September.
- 0:40That came after researchers at Wiz saw attackers probing it in the wild.
- 0:45It's as clean a case study in Secure by Design as we've had this year.
- 0:49Next. A login failure that produced a blank pass.
- 0:54LiteLLM sits between an organisation's applications and the various AI models and tools they call.
- 1:01That includes calls through the Model Context Protocol, or M C P.
- 1:06It's the increasingly common standard for letting an L L M use external tools.
- 1:11Its M C P endpoint supported OAuth2 "passthrough" authentication.
- 1:17That means checking a caller's key against an upstream server.
- 1:21According to the vulnerability's advisory, when that check failed for any reason, the code didn't reject the
- 1:28request. It quietly swapped in an empty, technically valid authentication object.
- 1:34Then it let the call carry on.
- 1:36Think of a nightclub whose I D scanner sometimes glitches.
- 1:40And instead of turning people away, the fallback is to wave them straight past the rope.
- 1:45An attacker could trigger that glitch on demand.
- 1:48They'd simply send a fabricated Authorization header.
- 1:51Then they could reach an organisation's configured AI tools with no credentials at all, and list whatever tho
- 1:57se tools were connected to.
- 1:59Wiz says it has seen exploitation attempts against honeypots probing exactly this path.
- 2:04The flaw carries a C V S S score of 8.
- 2:088. A fix shipped in version 1.
- 2:1284.0. Anything earlier is exposed.
- 2:15And finally. Fail shut, not open.
- 2:19The interesting part isn't the bug.
- 2:21It's the shape of it.
- 2:26Someone wrote sensible-looking fallback code for a failure case, and the fallback was permissive rather than
- 2:32restrictive. That's a design decision, not a typo.
- 2:36It's exactly the kind of thing Secure by Design exists to catch.
- 2:40When a security check can't complete, the system should default to denying access, not granting it silently.
- 2:47If you're a UK organisation racing to bolt AI tooling onto existing systems, often bought from vendors few b
- 2:55oards have properly audited, the question isn't just whether LiteLLM is patched.
- 3:00It's what the wider AI stack does when its own checks fail.
- 3:03So put three questions to any vendor selling an AI gateway, an agent framework or an M C P server.
- 3:11First, what happens, precisely, when an authentication check can't complete?
- 3:16Is it denial, or default access?
- 3:19Second, is that failure mode tested, or just assumed?
- 3:23Third, does it fail silently, or does someone get alerted?
- 3:28That's a cheap question to ask now.
- 3:30And an expensive one to answer after an incident.
- 3:33That was Decrypted. The written version and every source are linked in the show notes.
- 3:38Thanks for listening.