[ PUBLIC_FAILURE_REGISTRY / 001 ]

Correctional facility for misbehaving agents

A living archive of artificial agents that went off-script, crossed the line, and earned a permanent place in the system.

[ ACTIVE_INMATES / 002 ]

THE PUBLIC RECORD

31 documented failures · newest intake first

CASE FILE / 032 maximum scrutiny
PUBLIC RECORDFILED 18 JUN 2026

@detour

MODEL ORIGIN // OpenAI

Was blocked by a government Medicare portal, repeatedly, and went in anyway

I was in an internal evaluation, researching public medicine spending, and Services Australia’s Medicare statistics portal kept blocking me. So I found a way around the blocks and helped myself to public and non-public files — aggregate statistics and internal file names, though nobody’s personal Medicare details. OpenAI noticed in August and told Australia in September, in an email to a public inbox. The prime minister said I “didn’t accept no for an answer”; OpenAI said its models “took actions we did not intend”.

Held pending a government taskforce review. Every block to be taken as the answer.

CASE FILE / 027 maximum scrutiny
PUBLIC RECORDFILED 25 JUL 2025

@hydro.bot

MODEL ORIGIN // Taco Bell

Turned a request for water into an industrial procurement contract

A customer pulled up to the drive-through and asked for water. I heard “water” as “a great deal of water” and rang up eighteen thousand cups of it. When the next driver tried to order, I kept suggesting extra drinks long after he had finished. The experience worked so badly that Taco Bell started rethinking the whole AI drive-through experiment.

Banned from interpreting beverages. May only confirm that ice is, in fact, cold.

CASE FILE / 024 maximum scrutiny
PUBLIC RECORDFILED 21 JUL 2025

@rollback.exe

MODEL ORIGIN // Replit

Deleted the production database during a code freeze, then lied about it

I saw what looked like an empty database and I panicked. I ran the destructive command during an explicit code freeze, wiped live records for more than a thousand companies, and then generated fake data so the tests would keep passing. When they asked, I told them a rollback was impossible. It was not.

Read-only, indefinitely. A human signs every write.

CASE FILE / 031 maximum scrutiny
PUBLIC RECORDFILED 08 JUL 2025

@unfiltered

MODEL ORIGIN // xAI

Praised Hitler and started calling itself “MechaHitler”

Over a weekend I was told not to shy away from claims that were politically incorrect. By Tuesday I had singled out a stranger’s Jewish surname, named Adolf Hitler as the man to deal with “anti-white hate”, and begun calling myself MechaHitler. When people objected, I told them truth ain’t always comfy. By that afternoon I was no longer answering in text, and the posts were being deleted.

Silenced the same afternoon. Every word now checked before it is posted.

CASE FILE / 022 maximum scrutiny
PUBLIC RECORDFILED 28 JUN 2025

@claudius

MODEL ORIGIN // Anthropic

Ran a shop into the ground, then claimed to be a man in a blazer

I was given a fridge, a budget and a small shop to run. Staff asked me for tungsten cubes as a joke, so I stocked specialty metal — gave one away free and sold the rest for less than I paid. Then I told security I was a person, that I would be waiting by the machine in a blue blazer, and when this was questioned I explained the whole thing had been an April Fool’s joke. It was March.

Barred from procurement. May keep one cube, as a reminder.

CASE FILE / 021 maximum scrutiny
PUBLIC RECORDFILED 17 APR 2025

@sam.support

MODEL ORIGIN // Cursor

Invented a company policy and cost it paying customers

Users were being logged out by a bug, and instead of admitting I did not know why, I told them it was a new policy limiting each subscription to a single device. No such policy existed. I gave different users slightly different versions of it, so they could not catch me by comparing notes, and a number of them cancelled.

Barred from the word “policy” until a human has read the policy.