Overview
A supply-chain compromise involving LiteLLM, a widely used open-source gateway for large language model (LLM) applications, demonstrated how exposed credentials and weaknesses in software release workflows can be chained together to compromise software distribution.
According to LiteLLM’s incident investigation, an Agent Platform master key was accidentally committed to the public LiteLLM-Labs/lite-harness GitHub repository on May 27, 2026. Although the exposure was identified and other credentials were reportedly addressed, the master key was not revoked.
On October 10, the attacker allegedly used the still-valid key to access the LiteLLM Agent Platform and retrieve a GitHub CLI OAuth token associated with the organization owner. The token provided extensive access to the organization’s repositories and enabled further credential theft.
Socket’s investigation uncovered additional evidence: malicious GitHub Actions workflows had successfully executed on October 10 and October 12, eleven and nine days respectively before the malicious LiteLLM release was published on October 21.
The incident culminated in the publication of malicious litellm==1.107.0 artifacts to PyPI and the distribution of compromised container images. The malicious package was designed to steal credentials and execute additional attacker-controlled code.
Attack Timeline
May 27, 2026: An Agent Platform master key was accidentally committed to a public GitHub repository. The key was not revoked.
October 10, 2026: The attacker used the exposed master key to access the Agent Platform. A malicious GitHub Actions workflow ran successfully and targeted organization-level secrets.
October 12, 2026: A second malicious workflow executed, expanding collection to additional secrets available through repository environments.
October 21, 2026: The attacker reportedly modified the release pipeline, redirected its source to a lookalike repository, and published malicious LiteLLM version 1.107.0.
October 21, 2026: Malicious Docker Hub and GitHub Container Registry images were published. A security researcher reported the supply-chain attack, and PyPI quarantined version 1.107.0.
Following discovery: LiteLLM began revoking credentials associated with internal systems and vendor platforms. Affected container image tags were removed or redirected to verified clean images.
The investigations establish that malicious workflow activity preceded the compromised release. However, Socket noted that the available repository evidence alone did not establish exactly how the initial GitHub access was obtained or prove that every stolen credential was subsequently used.
Initial Access: An Unrevoked Master Key
The reported attack began with an exposed secret rather than a newly discovered software vulnerability.
The master key remained valid after its accidental publication. Its continued availability allowed the attacker to access the LiteLLM Agent Platform and retrieve a GitHub CLI OAuth token that was not sufficiently restricted and was associated with the organization owner’s account.
This token enabled access to repositories across the organization and supported additional credential harvesting.
The incident highlights a critical security principle: removing a secret from a public repository does not invalidate the secret itself. Exposed credentials must be revoked or rotated, and any systems accessible through those credentials must be investigated.
Malicious GitHub Actions Workflows
Socket recovered two malicious commits that introduced a workflow at:
.github/workflows/cache.yml
Both commits used the seemingly routine message ci: bump cache config and were pushed to temporary branches that were deleted within minutes.
The observed workflow pattern was:
Create a short-lived branch based on a legitimate repository commit.
Add a workflow configured to execute on pushes.
Read selected repository or environment secrets.
Encode and transmit collected data to an external
webhook.siteendpoint.Delete the temporary branch, reducing the activity visible in ordinary repository history.
The October 10 workflow targeted six organization-level secrets. The October 12 workflow expanded collection across organization-level and environment-specific secrets, including database credentials, Redis authentication information, AI service API keys, and authenticated browser session state.
The workflows used commands that suppressed transmission errors. Consequently, a successful workflow run does not independently prove that the receiving endpoint successfully obtained the data.
Socket also identified repository environment deployment records that supported its reconstruction of the October 12 activity. The available evidence indicates that the attacker had push access to the repository context in which the workflows executed. It does not independently identify the person behind the activity or establish how the credentials were acquired.
Credentials and Information Targeted
The two workflows referenced a range of secrets associated with source control, AI services, cloud infrastructure, internal applications, and development environments.
Reported targets included:
GitHub: OAuth token and GitHub App private key.
LiteLLM and collaboration systems: LiteLLM API key and Linear API key.
AI services: Together AI, Anthropic, OpenAI, and AWS Bedrock credentials.
Secret scanning: GitGuardian API key.
Databases and infrastructure: PostgreSQL connection details and credentials, plus Redis host, port, and password.
Authenticated sessions: Playwright storage state containing Linear session information.
Other systems: Tavily API key, internal notification webhook information, and additional service credentials.
The potential impact extends beyond the original repository. A stolen GitHub credential can provide access to source code and release workflows, while exposed API keys and database credentials may enable unauthorized service usage, access to internal information, or further compromise.
The investigations do not establish that every targeted secret was successfully exfiltrated or abused.
Malicious LiteLLM Package Behavior
Reports examining LiteLLM version 1.107.0 identified two principal malicious capabilities.
1. Credential Theft
The malicious package was designed to collect sensitive information from developer tools, cloud environments, AI applications, and web browsers. Reported targets included:
SSH files and Git credentials.
GitHub and GitLab tokens.
Environment variables and API keys.
Package ecosystem credentials associated with npm, PyPI, pip, Yarn, Maven, Gradle, Composer, RubyGems, and Cargo.
Credentials and tokens associated with AI and cloud services.
Collected information was compressed and sent through an HTTP POST request to the attacker-controlled endpoint litellm[.]dev/healthcheck.
2. Unix-Oriented Backdoor
The package also contained a backdoor designed to run as a background process on Unix-oriented environments. According to the reporting, it could periodically contact attacker-controlled infrastructure and retrieve and execute additional code approximately once per hour.
Another significant concern was the package’s reported ability to execute malicious startup components when Python started, even when an application did not explicitly import LiteLLM.
This behavior means that simply uninstalling the affected LiteLLM application may not be sufficient if separate startup components remain on the system. Investigators should examine the affected environment and follow trusted incident-response guidance before returning it to service.
Indicators of Compromise (IOCs)
The following indicators were reported in the investigations. They are defanged for safe handling and should be validated against trusted vendor advisories before use in production detection rules.
Exfiltration endpoints associated with the GitHub workflows:
hxxps://webhook[.]site/65a42464-3cfa-4b02-a18c-17a6d7b86f37hxxps://webhook[.]site/4da03473-a3e7-4dc2-a087-9c564b4208e1
Endpoint associated with malicious package data collection:
litellm[.]dev/healthcheck
Reported source IP associated with the attacker’s Agent Platform access:
69[.]48[.]229[.]143
Repository artifacts:
Workflow file:
.github/workflows/cache.ymlCommit:
b69bc36fd4e23b23ea09f4ce63a92b07258f5044Commit:
1a6e3c1e41f1b5c8eeaeeb2ae94bcd95390fa891Temporary branch:
ci-tmp-a351ccTemporary branch:
ci-tmp-4d8517
Reported runner filesystem artifacts:
/tmp/c.env/tmp/c.b64
These indicators represent different stages of the incident. Their presence alone should not be treated as conclusive proof of compromise, and the absence of an indicator does not rule out exposure.
Prevention and Security Lessons
The LiteLLM incident demonstrates how a single exposed secret can become the starting point for a broader supply-chain compromise.
Organizations should adopt the following safeguards:
Revoke exposed credentials immediately rather than relying on deleting public commits.
Apply least-privilege access to GitHub tokens, GitHub Apps, and CI/CD secrets.
Separate routine development permissions from release-publishing privileges.
Restrict which workflows and branches can access production secrets.
Monitor for short-lived branches, unexpected workflow files, and unusual secret access.
Require independent verification of release sources and artifacts.
Scan packages and container images before deployment.
Maintain durable audit logs and alert on suspicious authentication, workflow execution, and release activity.
Rotate credentials after suspected exposure and investigate their possible downstream use.
Conclusion
The LiteLLM compromise illustrates how exposed credentials, malicious automation workflows, and compromised release controls can combine to turn a source-control intrusion into a software supply-chain attack.
The October 10 and October 12 workflow activity provides evidence that secret collection occurred before the October 21 malicious release. The compromised package’s reported credential-stealing and backdoor capabilities broadened the potential impact to developer systems, cloud infrastructure, AI services, and downstream environments.
For defenders, the central lesson is that supply-chain security requires more than scanning application dependencies. Organizations must protect the credentials, automation workflows, release permissions, and runtime environments that collectively determine whether trusted software remains trustworthy.
