A poisoned LiteLLM release stole cloud and AI keys: check if you installed it
On 21 Oct 2026 an attacker published a malicious LiteLLM 1.107.0 to PyPI and Docker. It steals cloud credentials and AI provider keys. Here is who is affected and exactly what to do.
Updated 27 Oct 2026: see what changedHow urgent
Act now
Who is affected
Anyone who installed LiteLLM 1.107.0 from PyPI, or pulled the Docker tags v1.107.0, latest, main-latest or main-stable, on 21 Oct 2026
What to do
If you did, assume the machine is compromised: rebuild it from clean sources, pin LiteLLM to 1.106.2 or earlier, and rotate every secret it could reach
Free checklist
Published 23 Oct 2026 · Sources last checked 27 Oct 2026No. 10 / 87
Updates, follow-ups and corrections
- Update · 27 Oct 2026. LiteLLM published its root cause analysis on 23 Oct and says some earlier assumptions were inaccurate. According to the company: on 27 May a config file with the master key of an internal LiteLLM platform was committed to a public repository and the key was never revoked. On 10 Oct the attacker used it, and one API response contained the GitHub CLI token of an organisation owner. With it they pushed CI workflows that sent repository secrets, including two maintainers' tokens, to an external endpoint. On 21 Oct they pointed the release workflow at a lookalike fork, a stolen maintainer token approved the production step, and the packages went out around 11:40 UTC. The first token came from that API, not from logs as LiteLLM first suspected. LiteLLM still advises pinning 1.106.2 and says the Docker tags latest, main-latest and main-stable were repointed to clean images at 14:20 UTC. For a small team the lesson is simple: a key committed to a public repository stays valid until you revoke it, so deleting the file is not enough.
We keep the original text. When something changes, when we follow a story, or when we get something wrong, we add a dated note here instead of editing silently.
What happened
LiteLLM is an open-source tool that many AI projects use as a gateway to talk to different AI providers. On 21 October 2026 an attacker published a malicious version, 1.107.0, to PyPI (the Python package index) and to the project's container registries.
According to LiteLLM, the attacker got into its release process with stolen developer GitHub credentials, switched off protections and ran a release workflow that pointed to a look-alike repository. The malicious files were live on PyPI from about 11:39 to 13:23 UTC. Container images tagged v1.107.0, latest, main-latest and main-stable pointed to a compromised image from about 11:46 until about 14:20 UTC. Wiz estimates about 49,000 downloads of the bad version. LiteLLM says earlier versions are safe and that its private enterprise release was not affected.
What it does
According to Wiz, the package installs three files that run every time Python starts, even if your code never imports LiteLLM. They steal credentials and can load more code from the internet. The targets include:
- cloud credentials (AWS, Google Cloud, Azure);
- AI provider keys, such as OpenAI, Anthropic and Cohere;
- SSH and GPG keys, shell history and browser data;
- upstream provider keys stored in LiteLLM Proxy's database.
Wiz saw stolen AWS credentials being used about four and a half hours after they were taken. That is fast.
Who is affected
Anyone who installed 1.107.0, or ran a container built from the affected tags, during those windows. That includes laptops, build servers and the AI services you deploy. Automatic dependency updates can pull a bad release into several places before a person looks at it.
What to do
- Check what you ran. Look for LiteLLM 1.107.0 in your Python environments, your build logs, and the image tags you deploy. Version 1.106.2 and earlier are listed as safe.
- If you ran it, assume compromise. LiteLLM says finding no sign of an attack does not prove a system is clean.
- Do not run
pipfirst in an affected environment. Starting Python triggers the malwareMalware Any software made to do harm, such as stealing data, spying, locking files or giving an attacker control of a device. Tap again to open the glossary again. Remove the injected files (cook.pth,cook.py,sitecustomize.py), stop any leftover process, and rebuild the environment or image from clean sources, pinned to 1.106.2 or earlier. - Rotate every secretSecret (API key, token, password) Any piece of text that proves to a service that you are allowed in, such as an API key, a database password or an access token. Tap again to open the glossary that machine could reach, from a clean computer: cloud keys, AI provider keys, database passwords, SSH keys, repository tokens, and the environment secrets of any service it ran in. Our key leak video has the steps.
- Check your cloud logs for use of those credentials you do not recognise, and set spending limits and alerts.
- Pin your installs. Use exact versions and image digests, not tags such as
latest.
What we do not know yet
LiteLLM has not said how the attacker first got in. It suspects a recently patched vulnerabilityVulnerability A weakness in software, a setting or a process that an attacker could use to get in or do harm. Tap again to open the glossary in an internet-exposed internal deployment, and says some tools can leave GitHub tokens in logs, but none of this is confirmed. It also says it will publish more detail later. We will add a dated note here when it does.
If you build with AI tools
Our advice, not from the sources: if an AI tool or template set up your project, ask it which packages it installed, and keep that list. A package manager that installs the newest version automatically is convenient and risky. Waiting a few days before adopting new releases, and keeping secrets out of the machines that build your app, would have limited this kind of attack.
Terms explained in this article
- Supply chain attack: An attack that reaches you through something you already trust and install, such as a code package, a tool or a provider.
Watch the short video
The video loads from YouTube (privacy-enhanced mode) only when you press play.
Comments
No account needed. Add a nickname if you like, or leave it empty to comment anonymously. Comments are reviewed before they appear. Links are not allowed.
We store your comment, your nickname and a scrambled form of your connection that we use only to limit spam. We do not store your IP address or set cookies. Cloudflare Turnstile checks that you are not a bot.
Sources and verification
- LiteLLM: security incident of 21 Oct 2026 (vendor statement, status at 20:30 UTC) (opens in a new tab)
- Wiz: LiteLLM compromised, steals credentials (technical analysis, updated 22 Oct 2026) (opens in a new tab)
- Cyber Security News: LiteLLM compromised (22 Oct 2026, press report) (opens in a new tab)
- LiteLLM: the attack chain behind the v1.107.0 release (root cause analysis, 23 Oct 2026) (opens in a new tab)
We use primary sources first and say when we rely on a news report. Dates are the day we last checked each source.
Spotted a mistake? Email [email protected] and we will correct it and note the change.