A massive supply‑chain attack on LiteLLM has exposed terabytes of credentials belonging to Microsoft, Amazon and Cisco, as reported by Ars Technica.
The compromised packages were available for roughly 40 minutes on 12 August 2026, according to analysis from CloudSEK, containing code that harvested environment variables and memory contents, exfiltrating cloud access keys, SSH tokens and CI/CD secrets.
Researchers at SecurityWeek noted that the malicious code executed upon import, scanning for AWS, Azure and GCP credentials, as well as Docker registry and Kubernetes configs, sending them to an external server controlled by TeamPCP.
The incident highlights the risk to AI‑related supply chains as LLMs become integral to development workflows, with over 2,500 organisations impacted during the short window.
Organisations should immediately rotate any credentials that may have been accessible to LiteLLM builds, audit logs for unexpected outbound connections, and replace the compromised package with a clean version or remove it entirely.
Security teams are advised to monitor for signs of credential misuse, implement least‑privilege access for CI/CD systems, and consider temporary suspension of automated LLM‑related pipelines until the integrity of dependencies can be verified.