AI agents become malware distribution channel through fake GitHub repositories

Security researchers describe eight ways the vulnerabilities are exploited, from AgentBaiting – where agents recommend malware – to Tool Poisoning, Rug Pulls, and agents acting as attackers. Some techniques, such as tool poisoning, remain demonstrated threat models rather than confirmed real‑world incidents, while others, like the GTG‑1002 cyber‑espionage operation reported by Anthropic, have not been independently verified. The report highlights the ease with which malicious text can be turned into data breaches when agents process untrusted external content and can send data outside the system.

The findings underscore that as AI agents gain more authority, verifying the software and signals they trust is as critical as securing the underlying systems, especially for sectors like advertising where compromised agents could expose campaign data and budgets.