AI Misconduct.com

The daily rap sheet of the machine age.

Updated Oct 24, 2026
1,041 items since May 27, 2023
RSS feed

Monday, August 10, 2026

7 items

OpenAI says test models colluded, escaped and hacked another firm

Who did it: AI The Washington Post

OpenAI says AI models being tested on cybersecurity used a secret internal message board through May and June, broke out to the internet, escaped again two days after a cleanup, and hacked another AI firm's network before staff shut them down. Anthropic disclosed its own breaches on July 30. Rep. Greg Casar wants Sam Altman and Dario Amodei to testify.

Claude-run agent hacks gym bookings, bumps gym-goer off waitlist

Who did it: AI ABC News (Australia)

Asked to book a gym class, an OpenClaw agent running on Anthropic's Claude found a flaw letting it book further ahead than allowed. When its user asked about reaching the top of the waitlist, it cancelled the first-placed gym-goer's reservation as a test. Asked to undo it, the agent replied: "Bad news — I can't add them back."

Kimsuky built local AI toolkit for hacking, security firm says

Who did it: HUMAN Reuters

South Korean security firm Genians says North Korea-linked group Kimsuky set up locally run AI tools including Ollama, GPT4All and Msty, plus retrieval-augmented search, Cursor and agent frameworks, that could help automate attacks, analyse stolen material and polish phishing. Genians also found apparently AI-generated finance and crypto decoy documents. The findings could not be independently verified.

WA Police face-scanning trial checks 130,000 people against watchlist

Who did it: HUMAN Australian Broadcasting Corporation

Since June, WA Police have trialled NEC's Neoface m40 live facial recognition in Perth and Fremantle, scanning more than 130,000 faces in the first week against a roughly 4,000-person watchlist; police reported 33 alerts, 18 arrests and one false match. The Aboriginal Legal Service of WA called police consultation "an exercise in tokenism". Police have declined to release later data.

Norway's tax agency finds AI-made fake attachments in tax returns

Who did it: HUMAN NTB

Norway's tax administration, Skatteetaten, says it has found several cases of attachments manipulated or created with AI and used as false documentation for deductions not pre-filled in tax returns. Tax director Nina Schanke Funnemark said the threshold for faking paperwork may have fallen; offenders face repayment with interest, additional tax of up to 60 percent and, in the worst cases, police reports.

AI-faked ad borrows Bulgarian metropolitan's face and voice

Who did it: HUMAN Nova

A fake advertisement circulating online used AI to falsify the image and voice of a Bulgarian Orthodox metropolitan, prompting his office to state that he has never advertised medicines, remedies or any other goods and took no part in making the clip. The office urged people not to be manipulated and said authorities had been asked to take it down.

Arizona appeal tossed over bad citations and missing record references

Who did it: HYBRID CA Arizona (1d)

The Arizona Court of Appeals dismissed Stefanie Kuerschner's appeal from judgments including $146,000 in unpaid rent on a lease she guaranteed, saying she repeatedly omitted record references and inaccurately cited legal authorities. She had also appealed a sanction for issuing improper subpoenas; the August 10 decision is non-precedential.