Radware finds 145 million GitHub items affected by prompt injection
Researchers scanning 145 million GitHub items found that indirect prompt injection attacks on AI coding agents are rare but targeted, occurring at a rate of about 16 per million items. More than half of these attacks target maintainers' triage, review, or CI bots to steal secrets, hijack agents, or forge verdicts. Meanwhile, projects communicating with their own AI agents through hidden comments grew fivefold over four months, creating a new unauthenticated attack surface. Open article →