AI coding agents already look like attackers to EDR

Colleagues, I’d like to draw your attention to a cybersecurity development: AI coding agents are beginning to trigger endpoint protection rules that typically flag attackers.

Sophos has shown that Claude Code, Cursor and OpenAI Codex perform actions similar to malicious behaviour: accessing browser data, enumerating Windows credentials, downloading files via system utilities and writing to the Startup folder.

The issue is not malicious intent, but that behavioural engines are seeing the same signals.

My key takeaway: EDR rules need to be tied more precisely to the parent process, working directories and download sources, while access to credential stores should be tightly restricted.

Where do you see the line between useful automation and endpoint risk today?

#cybersecurity #EDR #AIAgents #ThreatDetection