Nederlands

LiteLLM supply chain attack: what to do if you installed 1.107.0

On 21 October 2026 a malicious LiteLLM release stole cloud and AI credentials. Find out if you are affected, clean the host and rotate the right keys.

Koen Dercksen4 min read

LiteLLM is a popular gateway between applications and model providers like OpenAI, Anthropic and Google. On 21 October 2026, attackers published a malicious LiteLLM release, version 1.107.0, to PyPI and Docker Hub. It stole cloud credentials and AI API keys from every machine that installed it. If your team runs LiteLLM, check your systems today.

Which versions and images are affected

The LiteLLM incident report lists these compromised artifacts:

Channel Affected Window on 21 October (UTC)
PyPI litellm==1.107.0, all seven wheels 11:39 to 13:23
Docker Hub tags v1.107.0, latest, main-latest, main-stable 11:46 to 14:20
GHCR the same four tags 11:46 to 13:51

According to LiteLLM, all earlier versions are safe. The official v1.107.0 source tag on GitHub doesn’t contain the malicious files, and enterprise customers on the private registry weren’t affected.

You’re at risk if an install or container start during that window used an unpinned version or one of these tags. CI pipelines and containers that pull latest are the most exposed.

What the malware does

Datadog Security Labs analysed the release. It drops three files into site-packages: cook.pth, cook.py and sitecustomize.py. Python runs .pth files and sitecustomize.py every time an interpreter starts, so every Python process in that environment runs the malware, even if it never imports LiteLLM.

Among other things, the credential stealer collects:

  • AWS, GCP and Azure credentials, including those from instance metadata.
  • Kubernetes, Docker, Vault and Terraform secrets.
  • SSH keys, Git, npm and PyPI tokens, .env files and shell history.
  • LLM provider API keys and sessions of developer AI tools such as Claude, Codex and Cursor.
  • The upstream provider credentials stored in the LiteLLM proxy database, if there is one.

A second file is a backdoor. Every hour, it asks a remote address for code to download and run. Datadog hadn’t seen a second-stage payload at the time of its analysis. Deleting the litellm package directory doesn’t remove the three files.

The attacker moved fast. Help Net Security cites research from Wiz: about four and a half hours after the theft, the attacker was testing AWS credentials taken from compromised CI runners and Kubernetes nodes, and checking access to Amazon Bedrock.

How the attackers got in

The attack chain report traces it all back to one key. On 27 May 2026, an admin key for an internal LiteLLM platform was committed to a public repository. Nobody revoked it, and four months later it still worked. With it, the attacker obtained a GitHub token belonging to an organisation owner, collected more tokens from there, and pointed the release pipeline at a lookalike repository.

This is LiteLLM’s second supply chain compromise of 2026. Datadog links an earlier incident in March, involving versions 1.82.7 and 1.82.8, to the TeamPCP campaign. When it published its report, Datadog had found no connection between the two attacks.

What to do now

These are LiteLLM’s steps. Follow them in this order.

  1. Don’t start with pip install. Every Python start in an affected environment runs the malware again, and that includes pip.
  2. Look for evidence. Search for litellm-1.107.0.dist-info, cook.pth and sitecustomize.py, and for containers that ran the image digest sha256:1bbcb1443bc962e45f5aad53fe3d2c30600e0379715fb6c664cdd3ac126c339e. Finding nothing doesn’t prove a host is clean.
  3. Remove the files from the shell. Delete cook.pth, cook.py, sitecustomize.py and their .pyc files from every affected site-packages, and stop the Python daemon holding /tmp/remote_code_loader.lock.
  4. Rebuild. Build fresh images and virtual environments from scratch, and pin litellm==1.106.2 or another explicit earlier version.
  5. Assume the theft succeeded on every host with outbound internet access.
  6. Rotate from a clean machine. That means the host’s IAM role, every provider key in the LiteLLM proxy database, the master and virtual keys, the database password and every secret in the host’s environment.
  7. Go through your cloud logs. In AWS CloudTrail, look for GetCallerIdentity calls around the install time, then for use of those credentials from outside your infrastructure.

And for the longer term:

  • Pin exact versions and image digests in production and CI. Don’t deploy latest.
  • Give the gateway its own provider keys with low spend limits, so a stolen key costs less.
  • Scan your public repositories for secrets and revoke any key that leaks. This whole attack started with a key nobody revoked.

How we approach this

An LLM gateway holds the keys to every model provider you use, which makes it a valuable target. In one project we built a private AI workspace with specialised agents, with an API layer between the agents and the model providers that logs usage and caches answers. We treat a layer like that as critical infrastructure and run it through the checklist above.

Sources

  1. Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised, LiteLLM, 21 October 2026
  2. The attack chain behind the v1.107.0 LiteLLM release, LiteLLM, 23 October 2026
  3. LiteLLM supply chain compromise: What you need to know, Datadog Security Labs, 21 October 2026
  4. LiteLLM hit by second supply chain attack this year, users urged to rotate credentials, Help Net Security, 22 October 2026

More articles