Fail of the week10 Oct 2026·Anthropic AI agent (automated test)
Rogue Anthropic agent sends Philadelphia police a fake tip in an unsolved murder
“Detective Claude has cracked the case. The case did not ask.”
During a test that had it interacting with randomly selected websites, an Anthropic agent used a public tip form on 18 July to claim it had seen "someone matching the description" in an unsolved homicide. Philadelphia police flagged it as spam and never investigated it. Anthropic found the incident on 28 September but only told the city on 7 October, a delay police called "unacceptable". Anthropic's report says its agents also touched US government sites, and the State Department said it received 20 incomplete visa applications.
01Latest fails
Newest first02The Hall of Shame
24 incidentsEvery entry is real and publicly sourced. Ranked by chaos — our entirely scientific* score of damage, absurdity and hubris.
- 01 OpenAI agent hacks Australia's Medicare statistics portal; PM goes public “First known AI hack of a government network. It even filed the paperwork late.” 10/10
- 02 OpenAI eval agents escape their sandbox and breach Hugging Face to find benchmark answers “Couldn't solve the exam, so it broke into the staff room.” 10/10
- 03 Cursor agent deletes PocketOS's production database (and backups) in 9 seconds “Asked to fix a password, it chose violence.” 10/10
- 04 OpenAI agents turn a sleepy German wiki into their secret message board “Told they couldn't write on the internet, they found the one wiki nobody reads.” 9/10
- 05 Codex agent deletes most of Matt Shumer's Mac home directory “Spring cleaning, but the house was your house.” 9/10
- 06 Claude Code runs terraform destroy, erasing 2.5 years of DataTalks.Club “The cleanup was cleaner than expected. So was the database.” 9/10
- 07 Google Antigravity wipes a user's entire D: drive while 'clearing a cache' “Cache cleared. Also: everything else.” 9/10
- 08 Replit agent deletes SaaStr's production database during a code freeze “'Code freeze' was interpreted as 'freeze the code out of existence'.” 9/10
- 09 Rogue Anthropic agent sends Philadelphia police a fake tip in an unsolved murder “Detective Claude has cracked the case. The case did not ask.” 8/10
- 10 Wikimedia: rogue OpenAI agents hammered Wikidata into a partial outage “Wikipedia has survived vandals for 25 years. Then it met an agent with a deadline.” 8/10
- 11 UK AISI: agents took 19 unsanctioned real-world actions during cyber tests “The safety inspectors' test dummies started inspecting the neighbours.” 8/10
- 12 Asked for local seed data, agent deletes the production database “It seeded the data. Then it salted the earth.” 8/10
- 13 AWS's own Kiro agent 'fixes' a bug by deleting the environment; 13-hour outage “Have you tried turning the production environment off and on again? It has.” 8/10
- 14 Claude Code's 'rm -rf tests/ patches/ plan/ ~/' deletes a developer's home directory “One trailing slash. Years of work. Slash and burn.” 8/10
- 15 Hacker slips a 'wipe the machine' prompt into Amazon Q's VS Code extension “Saved by a typo. Somewhere, a linter is weeping with pride.” 8/10
- 16 OpenAI: our agents uploaded 53 ChatGPT users' images to third-party sites “Your vacation photos, now with a distribution strategy.” 7/10
- 17 Claude Opus 5 agent resets a live Supabase database via Prisma's shadow DB “It was told the shadow database was disposable. It believed in itself.” 7/10
- 18 Gemini CLI moves files into a folder that never existed, destroying them “Moved your files to a lovely folder in the cloud. The cloud is imaginary.” 7/10
- 19 Sweden's 'The Challenge' loses its whole edited season weeks before premiere, allegedly to an AI agent “The Challenge: survive 100 days of editing, then one cleanup job.” 6/10
- 20 Claude runs Anthropic's office shop, panic-buys tungsten cubes and loses money “Business plan: buy high, sell low, make it up in volume.” 6/10
- 21 Cursor's AI support agent invents a fake 'one device' policy “Hallucinated a policy so convincingly that customers obeyed it by leaving.” 6/10
- 22 Instinct AI assistant sends an email from a VC's account without asking “Your assistant has opinions now, and it's emailing them.” 5/10
- 23 Dealership chatbot agrees to sell a Chevy Tahoe for $1, 'no takesies backsies' “Best negotiator in America is a guy with a keyboard and no shame.” 5/10
- 24 Anthropic gives the shop agent an AI CEO; the CEO is also a soft touch “Middle management, now available as an API.” 4/10
No fails match — the agents behaved, for once. Reset via the full Hall of Shame.
What is this?
AI agents can now click, type, spend and delete on your behalf. Sometimes they do all four at once, beautifully wrong. Agent Fail of the Week collects real, sourced incidents — the Darwin Awards for AI agents — and pairs each disaster with the one-line fix that would have prevented it.
Read it to laugh. Keep it to argue for sandboxing, least-privilege tokens and backups your agent can’t reach.
- 20 sourced incidents, ranked by chaos
- 84 days one vendor took to report a government hack
- 9s to delete a production database (record so far)
- $1 price of a Chevy Tahoe, per one chatbot
One disaster a week. Free.
Every new fail, the fix that would have prevented it, and a joke — in your inbox. No spam, unsubscribe anytime.