LiteLLM, the proxy that routes requests across many LLM providers, published version 1.107.0 on 21 October 2026 with an infostealer in every compiled wheel and a remote access trojan that loaded at every Python start.

Was I exposed? Check the filesystem, then the version, then the sentinels

Check the filesystem first, because starting Python in an affected environment re-runs the payload: no litellm-1.107.0.dist-info in site-packages, and pip show litellm should not report 1.107.0. LiteLLM’s incident timeline puts the seven poisoned wheels on PyPI from 11:39:51 UTC until PyPI quarantined the release at about 13:23 UTC on 21 October. Containers were exposed from 11:46:56 UTC through v1.107.0 and the floating latest, main-latest and main-stable tags. LiteLLM gives 14:20 UTC as the clean point for the Docker Hub tags; GHCR tags were unavailable by 13:51 UTC, so treat 14:20 as the vendor’s stated boundary. The sentinels show whether the payload ran: /tmp/.pc_once_72129118c59f5636c0bb means stolen data left the host, /tmp/remote_code_loader.lock means the daemon started. A missing marker proves nothing, since containers and cleanup remove them and Windows has no /tmp.

Three files that landed in the site-packages root

The payload was three files in the site-packages root, outside the litellm package directory. cook.pth is 12 bytes containing only import cook, so Python runs cook.py at every interpreter start, and cook.py decompresses and executes the stealer. sitecustomize.py also runs at interpreter start, launching a detached daemon that polls a GitHub dead drop hourly for a command address. The trojan runs only on Linux and macOS; on Windows it fails silently while cook.py still runs.

How the release was poisoned without touching the tag

A [tool.maturin] data = "cook" directive in pyproject.toml bundled a non-existent cook/ directory into every wheel; the payload is not in the tagged release but in fork commits above the tag and the build tree. LiteLLM’s incident post attributes the entry to stolen credentials, not a poisoned dependency: an internet-exposed internal deployment had logged Claude Code response bodies containing a user’s GitHub CLI OAuth token, and three developer accounts’ tokens came from those logs. One of those tokens belonged to a GitHub organisation owner and was used to disable branch protections and force-push a workflow change. The build then ran against the lookalike Berri-AI/litellm repository, and the pipeline signed and published the poisoned artifacts. An independent reconstruction from public build-provenance data, by Upwind’s research team, supports the decisive step: only v1.107.0 was built from the hyphenated repository.

What the stealer took

The collector also swept browser passwords, SSH and GPG keys, shell history, cloud and container credentials, CI/CD secrets and environment variables, then dumped the proxy database: LiteLLM_CredentialsTable, which stores provider API keys, and LiteLLM_ProxyModelTable.

AI coding-tool keys in the loot

The harvest also reads configuration and credential files for a long list of AI developer tools:

  • Claude Code
  • Claude Desktop
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Gemini CLI
  • Amazon Q
  • Codex CLI
  • Aider
  • Continue
  • HuggingFace

Sources differ on the exact membership beyond that overlap, so read the list as representative. Nothing in the payload treats AI credentials differently, so the breadth looks incidental rather than targeted.

One confirmed misuse, and a command channel that stayed dormant

The only observed misuse comes from Wiz’s analysis: AWS instance-role credentials checked with sts:GetCallerIdentity from two IPs between 16:20 and 17:01 UTC on 21 October, then Bedrock model listing, consistent with LLMjacking. Wiz also cites ClickPy data for roughly 49,000 downloads of the malicious version, a download count and not a victim count. The backdoor’s command channel was never armed: the GitHub dead drop held only test values and has been empty since 20 October 09:15 UTC. The exfiltration endpoint www[.]litellm[.]dev only receives data, so it is not a command server; it is a lookalike domain registered on 13 October 2026, unrelated to the project’s litellm.ai.

Rotate first, then pin

Any Python start in the affected environment re-runs the payload, including the pip you were about to use, so do not begin with pip. Delete cook.pth, cook.py and sitecustomize.py and their matching __pycache__ files in every affected environment, then kill the detached daemon holding /tmp/remote_code_loader.lock. Reinstall pinned to 1.106.2, rebuilding images and virtual environments from scratch where you can, then rotate the instance or task IAM role, every provider key in the proxy database plus the master and virtual keys, the DATABASE_URL password, and any secret in that host’s process environment.

What is still unverified

How the attacker reached LiteLLM’s internal deployment is unconfirmed, as is whether Claude Code routinely embeds GitHub CLI OAuth tokens in logged response bodies; that finding has no independent reproduction. No victim count or names have been published, and no cloud provider has confirmed the misuse Wiz observed. Cleanup beyond PyPI and Docker Hub is unaudited: mirrors, indexes, caches, pull-through registries and pulled layers were not checked. LiteLLM says the incident is contained, and no external forensic firm is named. One commit to the dead drop would re-arm every infected host.