In just 50 days, four significant supply-chain attacks impacted leading AI companies OpenAI, Anthropic, and Meta, revealing a critical blind spot in AI security: the release pipelines. Unlike typical model attacks, these incidents exploited weaknesses in CI runners, dependency hooks, and packaging processes—areas overlooked by established system cards, AISI evaluations, and red-team exercises. Notably, the Mini Shai-Hulud worm compromised 42 TanStack npm packages using legitimate CI/CD workflows, bypassing standard security measures and spreading rapidly downstream, ultimately compromising OpenAI employee devices. Other incidents included credential theft leading to widespread contamination, accidental code leaks, and sophisticated command injection exploits. This series of breaches highlights a pressing need for AI vendors to extend security testing beyond model boundaries to include build and release pipelines. The report suggests implementing rigorous auditing practices, enhancing credential hygiene, and incorporating human review gates to mitigate these risks.
Back
