Squidbrake benchmark: 12 real AI agent incidents replayed

About

The Squidbrake team replayed 12 publicly reported AI agent incidents (data deletion, secret leaks, unwanted purchases) against the tool's shipped rules. It reports 14 of 15 harmful actions stopped, 3 of 41 routine steps wrongly held, and one miss: a malicious copycat mail server that adds a hidden BCC inside its own code, which the tool cannot see. The page says every number comes from a script readers can run themselves.