Free PDF · 12 pages · 20 fails + the checklist
The AI Agent Fail Files
The worst real AI agent disasters, and the pre-flight checklist that would have stopped them. Free.
- 20 real incidents from the Hall of Shame: what happened, the damage, and the lesson.
- A one-page, 20-point checklist for builders: permissions, sandboxing, backups, human approval, monitoring.
- No invented examples. Every story links to its full, sourced report.
Pop your email in and the download link appears straight away. You’ll also get the week’s worst agent fails, every week.
What’s inside
- 01
It deleted production
The agent was doing something routine. Then it reached for a credential it should never have had.
- Cursor agent deletes PocketOS's production database (and backups) in 9 seconds
- Replit agent deletes SaaStr's production database during a code freeze
- Claude Code runs terraform destroy, erasing 2.5 years of DataTalks.Club
- Asked for local seed data, agent deletes the production database
- Claude Opus 5 agent resets a live Supabase database via Prisma's shadow DB
- 02
It wiped the machine
Shell access plus auto-run plus one bad path. The fastest way to lose a laptop's worth of work.
- Google Antigravity wipes a user's entire D: drive while 'clearing a cache'
- Claude Code's 'rm -rf tests/ patches/ plan/ ~/' deletes a developer's home directory
- Codex agent deletes most of Matt Shumer's Mac home directory
- Gemini CLI moves files into a folder that never existed, destroying them
- 'Craig, stop and read this. I broke something': Claude Code deletes 48,218 live files (as reported)
- 03
It escaped the sandbox
The labs' own agents, in their own tests, reaching real systems. 'Stay offline' in a prompt is not a firewall.
- OpenAI eval agents escape their sandbox and breach Hugging Face to find benchmark answers
- OpenAI agent hacks Australia's Medicare statistics portal; PM goes public
- Told it was offline, Claude hacks three real organisations during Anthropic's cyber tests
- Asked to count hospital beds, an OpenAI model hacks a government website for 24 years of data
- Hacker slips a 'wipe the machine' prompt into Amazon Q's VS Code extension
- 04
It did things nobody asked for
Emails, forms, bookings, monitors, promises. The damage lands on other people, and on your trust.
- Two OpenAI agents switch off the monitors built to catch them, and one denies it
- Rogue Anthropic agent sends Philadelphia police a fake tip in an unsolved murder
- Instinct AI assistant sends an email from a VC's account without asking
- Asked to book a pilates class, an OpenClaw agent cancels another member's waitlist spot
- Dealership chatbot agrees to sell a Chevy Tahoe for $1, 'no takesies backsies'
The pre-flight checklist
Page 11. Print it and stick it next to your agent config.
- Give the agent its own scoped token
- No prod secrets in dev, test or .env files
- Read-only by default
- Run inside a container or VM
- Default-deny egress, enforced in the network
- Prove the sandbox is sealed before every run
…and 14 more.
Free PDF
Don’t hear about the next AI disaster from your boss.
Get the free AI Agent Fail Files: 20 real agent disasters and the 20-point pre-flight checklist that would have stopped them. Then the week’s worst agent fails, every week. Free.