CyberKit

Threat intel / Malware and ransomware

LiteLLM 1.107.0: credential stealer shipped through the release pipeline

On 2026-10-21 a tampered LiteLLM 1.107.0 went out on PyPI and Docker for about two hours. Its payload stole developer and cloud credentials each time Python started.

High severity. Contained. High confidence. First reported 21 Oct 2026, last activity 23 Oct 2026. Briefing updated 31 Oct 2026.

AI-written summaries of public reporting, each checked against at least two sources from different publishers. Check the sources before you act. Last refreshed 2026-11-18 14:09 UTC. How this is made.

What is not settled: Neither source gives a victim count or the number of installs that pulled the release. How the first GitHub credentials were stolen is unknown, and no actor is named or attributed. The second-stage loader was not confirmed in the wild.

In plain words

What happened

Attackers published a poisoned version of LiteLLM, a popular open source tool for sending requests to AI models, to PyPI and Docker. It was available for about two hours. Anyone who installed or pulled version 1.107.0 got a program that runs whenever Python starts, even if LiteLLM is never used.

Who is behind it

No group is named. The attackers first used stolen developer GitHub credentials to run LiteLLM's release workflow from a lookalike repository, Berri-AI/litellm, which differs from the real one by a single hyphen. How the first credentials were stolen is not known.

What they want

Credentials. The program was built to collect the keys to developer accounts, cloud platforms and AI tools, which can open the way to other systems.

What they did

The release was published through LiteLLM's own pipeline, so it carried a valid build record. The code ran at Python startup and gathered configuration files, cloud and Git credentials, SSH and TLS keys, browser credential data, shell histories and environment variables. It then encrypted that material and sent it to an attacker-controlled server.

Who is at risk

Developers, build systems and AI gateways that installed 1.107.0 from PyPI, or pulled the Docker tags v1.107.0, latest, main-latest or main-stable, during the window. Whatever those machines could reach with their credentials is at risk too.

What to do

  • Find every copy of LiteLLM 1.107.0, including caches, internal mirrors, lock files, build outputs and container images, and check whether any of them ran during the window.
  • Remove the malicious startup files as well. Uninstalling the package alone does not remove them.
  • Move to 1.106.2 or an earlier version, then rotate every credential the affected machine could reach, working from a clean system.
The technical version, for hunting and patching

What happened

On 2026-10-21 the LiteLLM release pipeline published 1.107.0 to PyPI and to the GHCR and Docker Hub tags v1.107.0, latest, main-latest and main-stable. Datadog's timeline has the PyPI wheels live from 11:39 UTC, quarantined by 13:20, and the Docker images removed by 13:53. Microsoft's analysis, dated 2026-10-23, covers the wheels and images.

Who is behind it

Not established. Neither Microsoft nor Datadog names an actor, and neither ties this release to earlier LiteLLM activity. Microsoft says it has not determined how the attacker first gained access.

What they want

Secrets harvested from developer and build environments, which reach production clouds, CI/CD and internal registries.

What they did

Stolen GitHub CLI OAuth tokens from three engineers' accounts triggered the private release workflow against the fork. Per Datadog, the run was approved and the evidence of it removed. Per Microsoft, a credential tied to another developer account passed the production approval step. The wheels carried a .pth startup file, so Python runs the stealer on every interpreter start without importing litellm. Microsoft's analysis also describes a loader that could fetch and run more Python on Unix systems. The file it relied on was empty at analysis time, and no second stage was confirmed. Datadog saw a backdoor poll a remote source for instructions and found only test values there.

Who is at risk

Any environment where 1.107.0 was installed and Python was run, including developer machines, CI runners and shared gateways. The stealer fires at interpreter start, so an app that never loaded the package still counts. Microsoft lists cloud, Kubernetes, Docker, Git, SSH and CI/CD credentials among the targets.

What to do

  • Compare installed and cached LiteLLM artifacts against the published SHA-256 hashes, using a known-clean machine. Do not run pip or Python in the suspect environment to investigate it.
  • Review cloud and identity logs from the exposure window for use of the credentials you rotate, and revoke active sessions and tokens, not only passwords and keys.
  • Harden release pipelines: lock release workflows to the official repository, require approval from a separate account, and treat CLI OAuth tokens as high-value secrets.

ATT&CK techniques

Sectors: Technology, Cloud and hosting. Regions: Global.

Open next

Sources

Revision history

  • 2026-10-31: First published.

Newer: CVE-2026-110447: Nextcloud Collabora webshell attacks. Older: Capcom ID breach exposes 11.7 million account records.