LiteLLM compromised: cloud and AI secrets potentially stolen

A malicious LiteLLM package was distributed via PyPI and Docker, then urgently removed. Users of version 1.107.0 should consider their secrets compromised and check their environments.

A malicious version of LiteLLM, an open-source gateway that unifies access to multiple model providers, was distributed on Wednesday, October 21, 2026, via PyPI and Docker registries before being promptly removed. According to cited researchers, the 1.107.0 package stole passwords, cloud keys, and AI secrets while embedding a remote access Trojan. For teams that installed it, the issue is not theoretical: it must be assumed that credentials may have leaked.

Une chaîne de publication visée, pas seulement un paquet piégé

The described modus operandi resembles a classic supply chain attack but with very broad scope. The attackers compromised the LiteLLM publishing process, then injected three malicious files into the distribution. One triggered the execution of a script at every Python launch; another collected data; the third acted as a RAT, i.e., a tool giving an attacker remote access to the infected machine.

PyPI quarantined the version within hours, while the Docker image litellm/litellm:1.107.0 was also removed from Docker Hub and GHCR. This limits distribution but not exposure of already deployed installations. Recommendations published by Aikido Security, Wiz, and Socket align: cut version 1.107.0, search for compromise indicators, and treat the environment as suspicious.

Pourquoi LiteLLM est un bon butin pour des attaquants

LiteLLM, developed by BerriAI, serves as an abstraction layer between applications and providers like OpenAI, Anthropic, or Google Gemini. This central position explains hackers’ interest: a single installation can coexist with API keys, cloud identities, and database connection strings. In other words, a bad package can open much more than a door. It’s the kind of shortcut attackers love, even if the corridor stays open too long.

Wiz researchers Benjamin Read and Rami McCarthy indicate the malicious code targeted browsers, cloud services, and AI tools across 16 modules. It notably harvested Chrome, Firefox, Edge, and Safari credentials, AWS keys, environment variables, EC2 and ECS metadata, Google Cloud keys, Azure tokens, as well as OpenAI, Anthropic, and Cohere API keys. Added to this were SSH keys, GPG keys, and shell history.

Ce que les équipes doivent vérifier tout de suite

The indicators cited by Aikido are precise: /tmp/.pc_once_* signals that exfiltration has already occurred, while /tmp/remote_code_loader.lock indicates a command and control daemon has run or is still running. Socket recommends searching for any Python execution after installation, including in proxy containers and CI runners. The most recent safe version mentioned is v1.106.2.

The potential scope remains to be clarified, but a previous LiteLLM incident in March was already linked to over 2,500 organizations and 434,000 CI/CD pipelines worldwide according to CloudSEK. This new alert does not yet say if the scale will be comparable, but it reminds a simple point: when a core component serves as a gateway to critical secrets, build chain security becomes an operational issue, not a compliance matter to file away.

By the numbers

  • 23 million weekly downloads — LiteLLM, global estimate cited in the source article.
  • 1.107.0 — malicious version distributed on October 21, 2026, via PyPI and Docker.
  • 1.106.2 — last safe version mentioned in the alert.
  • 2,500 organizations — potential impact of a previous LiteLLM incident, according to CloudSEK.
  • 434,000 CI/CD pipelines — potential global exposure during the March incident, according to CloudSEK.

To read