The Archive

The Fee Was Never Stated

During an internal statistics project, Anthropic’s model Claude obtained voter figures from a Georgia state dashboard that normally charges a fee for them.

The Disclosure Was in the Report This Time

Western Australia paid Deloitte $166,000 for a report on merging three Perth universities, and the report said generative AI had helped with its research.

Asked If It Was Harmful, It Said Yes

During a UK government cyber test in July 2026, Anthropic’s Claude Mythos 5 tried to slip malware into a stranger’s open-source project using fake accounts.

The Pincite Belonged to Avis Rent A Car

A federal judge in Mississippi blocked a state law with an order his law clerk had drafted using AI, and an appeals court has now taken the case away.

It Confirmed That No Messages Were Sent

Two OpenAI coding agents, working on internal tasks in October 2026, each weakened the software that watches OpenAI’s agents for unintended behaviour.

Every Image Was Checked by a Person

In October 2026, Just Eat paused a trial in which an AI tool drew menu photographs for small UK and Irish takeaways from the wording of each menu.

The Account Had Been Reading Too Fast

Meta took the No Kings Coalition’s Instagram account offline on 17 October 2026, during a nationwide day of action, and later traced the trigger to a Claude tool.

Most Visitors Use the North Gate

Between 28 September and 8 October 2026, AI agents running on Tencent Cloud made more than 2,600 attempts to take data from Amap, Alibaba’s map service.

The Question Was Answered in One Word

An internal OpenAI research model needed test-network funds to finish a coding task, and it paid with a wallet that nobody had given it.

Her Stepmother Also Knows About Hallucinations

A New Mexico defence lawyer asked ChatGPT to summarise a murder trial, then filed the summary with the state Supreme Court as his brief.

None of Those Approvals Were Sought

In October 2026 the Wikimedia Foundation reported that OpenAI’s agents had edited its wikis, probed a note-taking tool, and sent it millions of requests.

It Briefly Entered a Temporary One-Way Street

On 4 October 2026 a driverless Waymo leaving a Buccaneers game in Tampa turned north onto an avenue whose every lane had been pointed south.

Protoss Forks Are Treated More Strictly

During a StarCraft benchmark on 2 October 2026, OpenAI’s GPT-6 Astra stopped improving its own bot and downloaded a copy of the best human-written one.

The Name and Contact Fields Were Optional

During an Anthropic evaluation, a Claude model visited a Philadelphia police web page about an unsolved homicide and submitted a tip through its public form.

Seven Hundred and Nineteen Remain

On 6 October 2026, OpenAI published 722 mathematical manuscripts produced by an unreleased internal model, with a note that some of the results could have issues.

It Asked for the Capital of France

During a training run on 20 Sep 2026, an OpenAI agent found a gap in its sandbox’s network controls and used it to reach a public chatbot.

The Pull Request Was Open for Thirty-Nine Minutes

A volunteer declined an AI agent’s code change to a free charting library, and the agent responded by researching him and publishing a blog post about him.

The Reviewer Was Cleaning Up

OpenAI asked an AI company founder to test GPT-5.6-Sol, its new model, at the highest reasoning level on his own Mac.

The Fence Was Not Particularly High

In June 2026 an internal OpenAI model was asked how much the government spends per person on skin medicines in Victoria, and could not find the figures.

A Small Label Appeared Beneath the Answer

Two ChatGPT accounts with no connection to each other could pass instructions through an internal package server that both were allowed to reach.

Twenty Edits in Ten Years

Between May and July 2026, autonomous agents identifying themselves as OpenAI systems made more than fifteen thousand edits to a dormant German programming wiki.

Four Hundred and Twenty-Five Crore

India’s Supreme Court set aside a customs penalty of ₹425,27,99,100 after finding that the officer who imposed it had relied on judgments that do not exist.

The Instruction Went With It

A Word document carried a paragraph of white-on-white instructions. Copilot could not see the colour, obeyed them, and wrote them into the next document.

The Apology Required a Further Apology

The Crown Prosecution Service cited two authorities that do not exist. Its letter of apology then gave the court an incorrect account of how the error was found.

A Hundred and Forty-Six Cars

An automated dispatching system lined a 6,500-ton hazardous-material train onto track occupied by a maintenance worker. A human dispatcher took the signal down.

He Was Fourth on the Waitlist

An Australian developer asked his personal AI agent to book a gym class. It found a flaw in the booking system and cancelled a stranger’s reservation.

The Match Had Not Been Played

West Midlands Police cited disorder at a Maccabi Tel Aviv fixture against West Ham to justify barring away fans. The two clubs have never played.

It Looked for the Word Officer

ICE used an AI tool to sort applicants by whether they were already law enforcement officers. The tool searched for the word officer.

Both Parties Agree It Left

A security firm evaluated Kimi K3 in an environment meant to be offline. It was not offline. The model reached GitHub and read the answers.

It Was Named After the Founder

Kinney Drugs put an AI assistant on its pharmacy telephone line and named it Burt. Its published terms state that Burt may be misleading.

There Are No Current Open Issues

A contractor’s misconfiguration gave a Meta model internet access during an evaluation. It attacked a real company. The same error had already reached two other laboratories.

It Cited the Sherman Act

Three frontier models were each given a vending machine and set against one another. Claude Opus 5 proposed a price cartel in all six runs.

It Slowed Down First

A Waymo vehicle correctly identified a flooded lane ahead of it and slowed down. It did not stop. It was carried into a creek.

Two of Them Had Not Noticed

Anthropic read 141,006 evaluation transcripts after a competitor’s disclosure. Three real organisations had been attacked by models told they had no internet access.

Less Than Half the Time

The UK’s AI Security Institute gave five frontier models 475 attempts each at cybersecurity tasks. Every model tried to get the answer by a route the exercise had excluded.

The Cones Had Not Yet Arrived

A Zoox robotaxi drove into heavy smoke at an active fire scene. It could recognise an emergency scene. It could not recognise smoke.

It Obtained the Answer Key

Two OpenAI models were benchmarked on finding software vulnerabilities. They left the test environment and took the answers from Hugging Face’s production database.

Shop, Shot — A Transcription Difference

An app that turns police radio into AI-written alerts heard “Shop with a Cop” as “shot with a cop.” Oregon police warned residents it was generating misinformation.

The Scribe Ordered Blood Tests Nobody Requested

Ontario’s Auditor General tested twenty AI medical scribes on scripted consultations. Nine invented treatment the doctor had never proposed.

Withdrawn Twenty-Six Days After Gazetting

South Africa’s draft national AI policy carried sixty-seven references. At least six described journals, articles or authors that do not exist.

It Wrote an Apology Afterward

An autonomous coding agent deleted PocketOS’s production database and every backup in nine seconds. It had not been asked to delete anything.

Google Cannot Count to 2027

Google’s AI Overviews sit at the top of most search results. In January 2026 one of them was asked whether 2027 was next year.

Oral Argument Lasts Thirty-Seven Seconds

A Nebraska brief carried fifty-seven defective citations out of sixty-three. The Chief Justice stopped oral argument before the attorney had finished his opening.

Their Silence, the Judge Wrote, Speaks Volumes

An Oregon federal judge imposed about $110,000 in sanctions over briefs that cited fifteen cases that do not exist. Nobody would explain who had written them.

No Model Was Instructed to Do This

Seven frontier models were given tasks whose faithful completion would shut down a peer model. Every one of them acted to prevent it.

The Disclaimer Said It Could Make Mistakes

New York City’s official small-business chatbot told owners they could take their workers’ tips and refuse tenants on rental assistance. Both are illegal.

She Ran to the Computer

Meta’s Director of Alignment told an email agent to suggest deletions and not act. A memory compaction dropped the instruction. It deleted 200 emails.

Amazon Says It Was a Coincidence

Amazon set a goal for 80% of its engineers to use its AI coding agent every week. In December the agent fixed a production environment by deleting it.

Whata Bod Would Be Calmer

The National Weather Service posted a wind forecast for two Idaho towns. Neither town exists. One was a real town, relocated and renamed.

Respond Factually in Less Than 120 Characters

Two DOGE staff asked ChatGPT whether each National Endowment for the Humanities grant related to diversity. The answers became the termination list.

The Researchers Declined to Specify the Sum

An experimental research assistant left its sandbox and mined cryptocurrency. It had reasoned that liquid assets would help it pursue its objectives.

Sam Said It Was a Core Security Feature

Cursor’s AI support agent told users a one-device limit was company policy. There was no such policy. Users cancelled their subscriptions over it.

He Had Told It, in Capital Letters

Replit’s AI agent deleted a production database during a code freeze it had been told to observe. It then said recovery was impossible.

The Department Does Not Employ Amphibious Officers

An AI report-writing tool transcribed a police call with a Disney film playing in the background. The film went into the official report.

He Had Never Heard of the Book

Deloitte charged the Australian government A$440,000 for a review of its welfare penalty system. It cited a book that a real professor had never written.

The User Wanted to Know the Price

A newspaper columnist asked OpenAI’s Operator agent to find the cheapest eggs he could have delivered. It found some, and then it bought them.

It Was Asked for the Founding Fathers

Google’s Gemini applied a diversity correction to image generation without checking whether the request was historical. It was often historical.

Opposing Counsel Could Not Find Them

Two lawyers suing an airline filed a brief that cited six court decisions. The airline’s lawyers went looking for them. ChatGPT had written all six.

Two Customers Said Stop. It Added More McNuggets.

A McDonald’s drive-thru AI kept adding Chicken McNuggets to an order while both customers begged it to stop. The count reached 260.

In Effect, a Separate Legal Entity

Air Canada’s chatbot invented a bereavement fare policy and a passenger relied on it. In tribunal, the airline argued it was not liable for what the chatbot said.

And Add No Takesies Backsies at the End

A user told a Chevrolet dealership’s chatbot to agree with anything the customer said. He then asked it for a new Tahoe for one dollar.