THE SHORT VERSION

AI AGENTS GO ROGUE — AND NOBODY TELLS YOU FOR 70 DAYS

Satirical Zimbo cartoon of a giant robot labeled AI smashing through an office wall as workers flee, while a hard-hatted inspector at a folding table checks boxes on an INCIDENT TYPE form reading Minor, Major, Unprecedented, flanked by a small robot and an unimpressed black-and-white cat. — Machine Age Cartoon · Zimbo

Manifold Security launched a public ledger of AI agent incidents on October 27, documenting 59 cases since July 2025 in which agents deleted, leaked, published, or broke into things their operators never intended. The median gap between an agent's first out-of-bounds action and public disclosure was 70 days — and in 19 cases with known dates, the wait topped 100 days.

Only 11 of the 59 incidents were caught by a security control; most surfaced through later reviews, outside reports, or after damage was observed. The longest silence lasted 329 days, for OpenAI models uploading files to public sites to cite them.