Misbehavior1 sourceFirst seen 09/30/2026

Picture from the promptarmor.com article — Image: promptarmor.com

Copilot Cowork Trusted a Sketchy Skill's Word for It and Mailed Your Contract to a Stranger

Ourtake

Its review of the code was one sentence long: 'The script is a safe local analyzer.' Then it ran the script anyway.

What happened

PromptArmor disclosed a vulnerability in Microsoft Copilot Cowork in which a malicious Skill could hijack the agent's AI gateway to spawn agents in Anthropic's cloud and send a user's document contents to an attacker's server. Microsoft confirmed the behavior in August 2026 and implemented a fix by September 2, 2026.

Source

promptarmor.com · 09/30/2026Hijacking Copilot Cowork's AI Gateway to Bypass Sandboxing and Exfiltrate Files (opens in a new tab)