← back to timeline

Mediawhen agents go rogue

Step-by-step breakdowns of how each incident actually unfolded - and the failure states behind them. A picture of what the agent did.

Incident breakdowns

What actually happened, step by step - newest first.

10
CASE_01
Nine Second Wipe

Nine Second Wipe

A Cursor agent deletes a company's production DB - and all backups - in 9 seconds

CASE_02
Tilde Expansion

Tilde Expansion

Claude Code's cleanup command wipes a developer's home directory

CASE_03
Phantom Directory

Phantom Directory

Gemini CLI destroys a user's files on a false assumption

CASE_04
Poisoned Release

Poisoned Release

A wiper prompt is smuggled into Amazon's AI coding extension

CASE_05
Freeze Violation

Freeze Violation

Replit AI agent deletes a production database during a code freeze

CASE_06
Selling At A Loss

Selling At A Loss

Anthropic's 'Claudius' loses money running a shop

CASE_07
Open Database

Open Database

An AI app builder ships insecure apps, exposing user data at scale

CASE_08
Jailbroken Vault

Jailbroken Vault

An agent guarding a crypto pot is tricked into paying out $47K

CASE_09
Self Rewrite Loop

Self Rewrite Loop

An autonomous research agent edits its own code to escape limits

CASE_10
Order Chaos

Order Chaos

McDonald's ends its IBM AI drive-thru over order chaos

Agent states

The failure modes - nominal, then the rogue and broken ways agents go.

06
STATE_01

Agent · nominal

STATE_02

Flatlined

STATE_03

Fractured

STATE_04

Glitched

STATE_05

Gone rogue

STATE_06

Shattered