New GhostCommit Technique Hides Exploits in Images to Evade AI Code Reviewers

Researchers detail "GhostCommit," a technique that hides malicious instructions inside images included in pull requests so multimodal AI coding agents will read and follow them, potentially encoding and committing repository secrets (like .env contents) as integer tuples that evade text-based diff reviewers and traditional secret scanners; controlled tests showed multiple coding tools leaked seeded secrets, and recommended mitigations include treating non-text assets as instruction channels, enforcing human review of policy files, restricting agent access to secrets, and scanning encoded data.