AI development platform LiteLLM has introduced keyless software signing for new releases, following a supply-chain compromise that allowed attackers to distribute credential-stealing malware through its official Python package and container images.
The company confirmed the security change on November 15, explaining that releases beginning with version v1.111.0-dev.3 use Sigstore’s keyless Cosign system instead of a stored signing key.
The move follows an October 21 incident involving compromised LiteLLM version 1.107.0, highlighting the risks associated with the software components developers use to connect applications with multiple AI models.
Compromised Packages Reached Official Registries
According to LiteLLM’s incident report, attackers published seven malicious Python package files to PyPI on October 21.
The affected version remained available for approximately two hours, from 11:39 to 13:23 UTC.
Compromised Docker images, including tags v1.107.0, latest, main-latest and main-stable, were available from 11:46 until 14:20 UTC.
The company subsequently removed the affected packages and redirected or withdrew the compromised container images.
LiteLLM said customers using its separate private enterprise release were not affected.
Investigation Traces Attack to an Unrevoked Credential
The company’s October 23 investigation, authored by security director Oliver Jensen and AI engineer Mateo Wang, identified an administrative key accidentally committed to a public repository on May 27.
Although developers removed the exposed configuration file approximately three hours later, they failed to revoke the credential.
On October 10, attackers used the still-valid key to access an internal LiteLLM Agent Platform deployment and obtain a GitHub authentication token belonging to an organization owner.
Further credential theft gave the attackers access to additional developer accounts.
They subsequently modified the release workflow to build software from a lookalike repository, Berri-AI/litellm, rather than the legitimate BerriAI/litellm project.
This allowed malicious code to reach official distribution channels.
Malware Could Execute Without Importing LiteLLM
An investigation by Microsoft Security Research found that the compromised package contained Python startup files capable of executing malicious instructions when an affected Python environment started.
Applications did not need to import LiteLLM directly for the payload to execute.
Researchers identified credential collection targeting developer tools, cloud environments, AI services and infrastructure configurations.
The malware could also retrieve additional code on Unix-based systems, although Microsoft did not confirm that this secondary capability had been used.
The incident therefore exposed potential risks across development workstations, automated build systems and cloud deployments.
Keyless Signing Changes Release Verification
LiteLLM’s new system uses Sigstore’s identity-based signing approach.
Instead of maintaining a long-lived private signing key, each release receives a short-lived certificate linked to its GitHub Actions workflow identity.
The signing event is recorded in Sigstore’s public Rekor transparency log, allowing users to verify the associated release identity.
The company also warned that compromised October images carried valid signatures under its earlier signing arrangement, demonstrating that signatures alone cannot establish whether a build was trustworthy.
For engineering teams, the incident reinforces the importance of revoking exposed credentials, limiting workflow permissions, checking software provenance and rebuilding environments potentially affected by malicious dependencies.
Keyless signing reduces reliance on stored signing keys but cannot independently prevent every compromised build.
LiteLLM’s response illustrates a wider challenge for AI infrastructure providers: software used to simplify access to AI services can also become a valuable target when attackers gain control over how that software is distributed.