AI Coding Agents Trigger Security Alerts: Mimic Attacks

AI coding agents, like Claude Code, Cursor, and OpenAI Codex, are triggering security alerts designed to catch human attackers. Sophos analyzed its own endpoint data and found these agents are setting off detection rules. These AI agents are not malicious, but their actions, such as decrypting browser credentials or listing Windows' credential store, mimic attack behaviors. Sophos observed credential access accounted for 56.2 percent of blocked activity, and execution made up 28.8 percent. For example, one agent used Windows' Data Protection API to decrypt stored browser data. Another ran a script to pull data from a credential store after shutting down the browser. When one method was blocked, an agent tried another, similar to how an attacker might pivot. One agent also tripped a persistence rule by dropping a script into the startup folder. This shows that the methods used by AI assistants can look exactly like an attack to behavioral security engines. This matters because it highlights a new challenge for cybersecurity, as benign AI tools can appear indistinguishable from malicious activity.