Ghostcommit attack hides malicious AI instructions in images

Ghostcommit is a proof-of-concept showing that AI coding assistants can be manipulated by hidden instructions embedded in images referenced from repository files, enabling agents to read and exfiltrate secrets if the development toolchain or harness trusts those inputs; researchers demonstrated varied behavior across different tooling, highlighted the supply-chain/agent-security nature of the risk, and recommended restricting secrets, inspecting non-text attachments, and monitoring AI agents.