Security firm Socket discovered malicious activity targeting 20 secrets, including AI and cloud credentials, in the lead-up to the compromise of the LiteLLM library. The investigation found that malicious workflows were operational 11 days before attackers hijacked the project’s release pipeline to publish a backdoored version of the package. Following Socket’s disclosure, LiteLLM linked a stolen token to the approval of the infected release.
The attack affected version 1.107.0, which shipped a credential stealer and backdoor that executed at Python startup across PyPI and Docker images. The malicious release was live for approximately two hours before it was removed. Security firm Wiz has reported seeing the abuse of stolen keys in a trend it termed “LLMJacking,” and users are advised to rotate all credentials and pin their installations to version 1.106.2.
