A Poisoned Security Scanner Just Backdoored the AI Stack
On March 24, a threat actor called TeamPCP published backdoored versions of LiteLLM (versions 1.82.7 and 1.82.8), a widely used Python package for AI model routing downloaded 3.4 million times per day. They got in by first compromising Trivy, Aqua Security's open source vulnerability scanner, on March 19. The poisoned Trivy GitHub Action exfiltrated PyPI credentials from LiteLLM's CI/CD pipeline.
The malicious versions contained a three-stage payload: credential harvester (SSH keys, cloud creds, Kubernetes secrets, crypto wallets, .env files), Kubernetes lateral movement deploying privileged pods to every node, and a persistent systemd backdoor polling checkmarx[.]zone for more binaries. The packages were live for about 3 hours before PyPI quarantined them. Version 1.82.8 was especially nasty, using a .pth file that fires on every Python startup, not just LiteLLM imports. It caused an accidental fork bomb that's how a researcher at FutureSearch discovered it.
Microsoft published remediation guidance. The attacker even tried to close the disclosure issue on GitHub using compromised maintainer credentials and flooded it with bot comments. TeamPCP has now hit five ecosystems: GitHub Actions, Docker Hub, npm, Open VSX, and PyPI. They posted on Telegram bragging about "stealing terabytes of trade secrets" and said they're partnering with other groups.
As Gal Nagli at Wiz put it: "The open source supply chain is collapsing in on itself. Trivy gets compromised, LiteLLM gets compromised, credentials from tens of thousands of environments end up in attacker hands, and those credentials lead to the next compromise. We are stuck in a loop."
We are in a new wild west with AI technology. The rush to ship AI products has created an enormous attack surface that most companies are not prepared to defend. LiteLLM is downloaded 3.4 million times per day. Trivy is used by thousands of organizations to secure their infrastructure. When tools this embedded in the stack get compromised, the blast radius is measured in tens of thousands of environments, not individual machines.
Security best practices need to improve dramatically across the entire software stack. Pin your dependencies. Verify your CI/CD tooling. Audit your supply chain. The era of blindly trusting open source packages because they have a recognizable name is over. The AI boom is moving faster than the security infrastructure designed to protect it, and attackers like TeamPCP are exploiting that gap with precision. The tools we trust to find vulnerabilities are now introducing them. Until the industry treats supply chain security with the same urgency it treats model performance, we are stuck in a loop.
|