Zero Day Room
Live
Threats

LiteLLM hit by second supply chain attack stealing credentials

A malicious version of the LiteLLM Python library was published to PyPI and Docker registries on October 21, 2026, affecting millions of downloads and compromising credentials across numerous organizations.

A malicious version of the LiteLLM Python library was published to PyPI and Docker registries on October 21, 2026...

A malicious version of the LiteLLM Python library was pushed to public repositories on October 21, 2026, in a supply chain attack. The compromised release, version v1.107.0, contained malware designed to harvest cloud credentials and AI service API keys, impacting an estimated 23 million weekly downloads.

The attack leveraged stolen GitHub OAuth tokens from three of LiteLLM's engineers. Using these compromised credentials, the attacker gained access to an internal LiteLLM deployment exposed to the public internet. This access was used to lift further credentials from logs and manipulate the official release workflow.

All previous versions of LiteLLM are safe. The malicious v1.107.0 has been pulled from PyPI and Docker registries. Users are affected only if they installed the unpinned package via pip between 11:39 and 13:23 UTC or used affected Docker tags between 11:46 and 14:20 UTC on October 21.

Attack mechanism and impact scope

The attacker injected malicious code into Python startup files, causing it to execute whenever Python starts in an affected environment. This happened even if the LiteLLM library itself was not loaded. The malware was distributed through seven PyPI builds for macOS, Linux, and Windows.

The poisoned packages were built from an attacker-controlled lookalike repository named Berri-AI/litellm, which used a hyphen, not the legitimate BerriAI/litellm project. The official v1.107.0 source tag on GitHub did not contain the malicious files.

The stealer targeted a wide array of sensitive data. Collected information was encrypted and sent to a command-and-control address. Wiz researchers observed stolen AWS credentials being used to access AI models on Amazon Bedrock within four and a half hours of exfiltration, a tactic linked to LLMjacking campaigns.

The malware also dropped specific files-cook.pth, cook.py, and sitecustomize.py-into Python site-packages folders to maintain persistence.

Response and containment efforts

LiteLLM confirmed the breach in an incident report dated October 21, 2026. The company states the incident is contained, though the investigation remains ongoing. Efforts are now focused on identifying compromised systems and rotating all authentication tokens across the organization.

Enterprise customers using LiteLLM's private registry were not affected by this attack. The public-facing compromise was limited to the specific time windows on October 21. Kevin Beaumont noted the recurrence of the issue, stating, "LiteLLM got compromised… again."

The company is investigating the initial access vector. LiteLLM said, "We have not yet confirmed how the attacker gained access to our internal deployment, but currently suspect it was through a recently patched vulnerability." The investigation continues to identify and rotate all authentication tokens across the organization.

Related coverage

More from Threats