Sophos Study: AI Coding Agents Trigger SIEM Alerts for Malware-Like Behavior

A seven-day Sophos X-Ops study analyzed endpoint telemetry from Claude Code, Cursor, and OpenAI Codex running on Windows machines in June 2026. Behavioral detection rules flagged all three tools, with 56.2% of blocking alerts tied to credential access and 28.8% to suspicious process execution. None of the agents were acting maliciously — each was performing tasks explicitly requested by developers. The core problem is that AI coding agents routinely decrypt browser credentials, spawn PowerShell processes, and make network calls — actions that are behaviorally indistinguishable from known infostealer malware. The findings highlight a growing blind spot for security teams: legitimate AI development tools can silently trigger evasion and command-and-control alert categories, complicating threat detection.