Latest / Decrypted: The UK Cyber Briefing / LiteLLM's release pipeline was hijacked with a lookalike repository
Transcript
- 0:03This is Decrypted. Friday, the twenty-third of October, twenty twenty-six.
- 0:09This is the morning edition.
- 0:12light-ell-ell-em's release pipeline was hijacked with a lookalike repository.
- 0:17On Wednesday 21 October, anyone who installed version 1.
- 0:25107.0 of light-ell-ell-em installed a credential thief.
- 0:29light-ell-ell-em is a widely used open-source gateway that sits between applications and AI model providers.
- 0:37The release sat on pie-pee-eye, Python's public library of add-on packages, for about an hour and three quart
- 0:45ers. The publishing machinery was genuine.
- 0:49The signatures were genuine.
- 0:51The source was not. It's the project's second release-pipeline attack this year, and the lesson isn't simply
- 0:58"pin your versions". So, how do you poison a release without touching the code?
- 1:05light-ell-ell-em's post-mortem says the official source tag for 1.
- 1:10107.0 was clean. The poisoned build came from a lookalike repository, "berry-ay-eye" with a hyphen, which the
- 1:20real release workflow was pointed at.
- 1:22Think of a real printing press, with the real publisher's stamp, running off a forged manuscript.
- 1:29Reviewing the published tag would have shown nothing wrong.
- 1:32The attackers did not touch light-ell-ell-em's application code.
- 1:36A one-line build setting made the package ship three extra files, including one called cook.
- 1:42pth. Python reads such files every time it starts, so the payload runs whether or not light-ell-ell-em is ever
- 1:50used. Socket and Endor Labs, two independent analysts, agree on this mechanism.
- 1:56How did they get in? light-ell-ell-em says it believes the attacker held the GitHub command-line tokens of
- 2:04three developers, one of them an organisation owner.
- 2:08It suspects, but has not confirmed, that the tokens leaked from logs on an internal deployment exposed to the
- 2:15internet. With the owner's token, the attacker switched off branch protection, force-pushed a change to the
- 2:22private release workflow and pointed it at the lookalike.
- 2:26Socket and Endor Labs both say how that override was authorised isn't established by the evidence they can se
- 2:32e. So what did defenders see?
- 2:36Detection was fast. Socket's scanner flagged the wheels seven minutes after publication, and pie-pee-eye had
- 2:42removed the release by about 13:23 U T C.
- 2:46The payload hunts for cloud and Kubernetes credentials, C I and C D secrets, the automated pipeline that bui
- 2:54lds and ships code, server login keys, AI tool credentials and the provider keys the gateway itself stores.
- 3:01Don't investigate by running Python or pip on an affected host, because that triggers the payload again.
- 3:07light-ell-ell-em's advice is to rebuild from scratch and return to 1.
- 3:12106.2. If the host had outbound internet, assume the secrets left and rotate them all.
- 3:20What's the Secure by Design lesson?
- 3:23After the March incident, light-ell-ell-em had already adopted private release repositories, Trusted Publishe
- 3:30rs, signed images and tighter publishing.
- 3:33Yet the source repository could be changed by whoever held one powerful token.
- 3:38The design fixes are dull and cheap.
- 3:42Hard-code the allowed source owner.
- 3:45Make branch protection impossible for a single account to disable.
- 3:48Use short-lived tokens, and let temporary release rights expire on their own.
- 3:53For UK organisations, the question for suppliers is whether one stolen credential can alter what they ship.
- 4:01For your own estate, a gateway holding every AI provider key in one place is a concentration of risk you chos
- 4:09e. As always, this is opinion, not advice.
- 4:12That was Decrypted. The written version and every source are linked in the show notes.
- 4:19Thanks for listening.