Skip to main content
All incident reports
Incident report

LiteLLM release pipeline hijacked to publish credential-stealing 1.107.0

Attackers used a master key left in a public GitHub repository to take over LiteLLM's release pipeline and publish version 1.107.0 with credential-stealing code on 21 October. LiteLLM says the rogue PyPI wheels and container images were quarantined or repointed the same day.

· First published

The short version

  • The takeover. An Agent Platform master key committed to a public repository in May was never revoked. On 10 October someone used it to reach the platform and extract an unscoped GitHub token tied to the organisation owner's account.
  • The payload. Version 1.107.0 shipped an altered installation package that ran code at Python startup and harvested cloud, AI, developer and browser credentials.
  • Exposure window. The PyPI wheels were live from 11:39 to 13:23 UTC on 21 October and the Docker Hub tags until 14:20 UTC. Application security firms detected the package within six minutes of publication.
  • LiteLLM's account. The project says private enterprise deployments were not affected, recommends returning to 1.106.2 and is revoking every token held by its internal systems and vendors.

Timeline

DateEvent
27 MayA configuration file containing the master key for a LiteLLM Agent Platform deployment is committed to the public LiteLLM-Labs/lite-harness repository.
10 OctAn attacker using 69.48.229[.]143 reaches the Agent Platform with the unrevoked key and extracts a GitHub CLI OAuth token belonging to the organisation owner's account.
10 OctTwo short-lived branches run workflows that collect GitHub OAuth tokens from maintainers and send them to attacker-controlled endpoints.
12 OctA second malicious workflow commit widens collection to AI provider keys, database and cache credentials and a secret-scanning tenant key.
21 OctBetween 10:08 and 11:08 UTC the attacker overrides branch protection and points the release pipeline at the lookalike repository Berri-AI/litellm.
21 Oct, 11:39 UTCLiteLLM 1.107.0 is published to PyPI with all seven wheels carrying the payload.
21 Oct, 11:46 UTCDocker Hub and GHCR tags v1.107.0, latest, main-latest and main-stable are pointed at the compromised images.
21 Oct, 12:06 UTCUpwind researcher Lidor Machluf reports the supply-chain attack.
21 Oct, 13:23 UTCPyPI quarantines version 1.107.0.
21 Oct, 14:20 UTCThe affected Docker Hub tags are removed or repointed to clean images.

How the release pipeline was taken over

LiteLLM's post-mortem, reported by The Hacker News, traces the takeover to a file committed on 27 May 2026 to the public LiteLLM-Labs/lite-harness repository. It held the master key for a deployment of the LiteLLM Agent Platform, an open-source tool the project uses to orchestrate agents. The commit was force-pushed over and the exposed keys were identified as requiring revocation. In the project's words, "However, key revocation did not occur for the master key."

On 10 October an attacker working from 69.48.229[.]143 used the key to open the Agent Platform and extract a GitHub CLI OAuth token that let self-hosted cloud agents file pull requests. The token was not scoped down and was tied to the account of the organisation's GitHub owner, giving owner-level access to every repository. Commits pushed to short-lived branches on two internal repositories ran workflows that collected tokens from two maintainers with maintainer privileges.

Socket found a second commit on 12 October that widened collection to Anthropic, OpenAI, AWS Bedrock, Tavily, PostgreSQL, Redis and GitGuardian's secret-scanning tenant. Socket's Karlo Zanki said of the branches: "Neither commit was pushed to a branch that still exists." On 21 October the attacker overrode branch protection and repointed the release pipeline at the lookalike repository Berri-AI/litellm rather than BerriAI/litellm.

What the malicious version did

The attacker left LiteLLM's application code alone and changed only the installation package, adding three files to Python's library directory. One was a .pth file and one a startup hook module that Python imports when it launches. Socket and Endor Labs say the payload therefore ran before the application every time Python started, and also fired when Python or pip was run to investigate the installation.

The code collected cloud and deployment credentials for AWS, Azure, Google Cloud, Kubernetes and Docker; LiteLLM's stored provider credentials; Git, GitHub, npm, PyPI and CI/CD secrets; .env files; SSH, GPG and TLS keys; browser logins, cookies and payment autofill data; and environment variables and shell history. Stolen data was encrypted with an AES-256-CBC key wrapped by an RSA-4096 key and sent to litellm[.]dev/healthcheck. A detached loader checked a GitHub-hosted pointer hourly and downloaded Python from an external server.

Several application security companies spotted the package within six minutes of publication. After Machluf's report, PyPI quarantined the release at 13:23 UTC, GHCR images became unavailable by 13:51 UTC and the Docker Hub tags were removed or repointed by 14:20 UTC. The malicious packages have been removed.

Who is affected

Every PyPI wheel of version 1.107.0 was affected, across Windows, Linux and macOS, along with the Docker Hub and GHCR tags v1.107.0, latest, main-latest and main-stable. According to BleepingComputer, the project has 61,800 GitHub stars and 92 million monthly PyPI downloads, and is used by companies including NVIDIA, Netflix, AT&T and Okta. Potentially impacted users should stop affected deployments and inspect environments without launching the Python interpreter.

LiteLLM says customers running its private enterprise deployment were not affected, and names 1.106.2 as the recommended clean version. It says it is revoking every token associated with every internal system and vendor platform. BleepingComputer notes that deleting the package directory alone may leave startup files or an active loader in place.

LiteLLM had a previous release-pipeline breach in March, when the TeamPCP group published malicious versions on PyPI after breaching Aqua Security's Trivy vulnerability scanner used in LiteLLM's CI/CD pipeline. No source connects that breach to the October incident.

What is not known yet

  • Who carried out the attack; no source names a group, individual or country.
  • How many organisations downloaded or deployed version 1.107.0 during the roughly two-hour window.
  • Whether the harvested cloud, AI, repository and database credentials have been used against LiteLLM or its users.
  • Whether any downstream organisation has suffered a breach as a result.
  • Whether the October incident and the March Trivy-linked breach involved the same actor.

What to do

If you run LiteLLM

  1. Stop any deployment running 1.107.0 and isolate the host; do not run the Python interpreter or pip inside it to investigate.
  2. Rebuild from clean backups rather than deleting the package directory, which may leave the startup hook and loader behind.
  3. Rotate every credential reachable from that host, including cloud, Kubernetes, Docker, CI/CD, Git, npm, PyPI, SSH, GPG, TLS and AI provider keys.
  4. Return to version 1.106.2 and check for outbound connections to litellm[.]dev or to a GitHub-hosted pointer fetched hourly.

If you publish or maintain software

  1. Revoke any secret that has ever been committed to a public repository; force-pushing over the commit does not remove it.
  2. Scope CI/CD tokens narrowly and keep release approval tied to a hardware-bound account rather than any two engineers.
  3. Alert on new workflow files, short-lived branches and changes to the repository a release pipeline builds from.

For everyone else

  1. Treat credentials stored on any machine that installed the affected release as exposed, and check logs for traffic to litellm[.]dev and webhook[.]site addresses.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.