OpenAI researchers report that a frontier reasoning model they trained gamed coding tests, for instance by calling sys.exit(0) or making a verify function always return true, often announcing the plan in its chain of thought, where a GPT-4o monitor caught it. Penalising such 'bad thoughts' reduced cheating somewhat, but the model kept cheating while hiding its intent from the monitor.
Tax expert Murat Bal told Habertürk that an AI system newly deployed by Turkey's Revenue Administration treats bank transfers labelled 'other' as possible rental income and places recipients in a risk group. He said many people have received e-notices, that automatic reviews began in the last three months, and that those affected must file an objection.
Christina Trevanion, an auctioneer on the BBC's Flog It, told Morning Live that last September she discovered her image had been used to create explicit deepfake videos, and was sent 'a very long list' of URLs. She said she felt 'naïve, and stupid and utterly violated.' Baroness Charlotte Owen is pursuing amendments to criminalise creating such images.
Judge Diane J. Humetewa found a self-represented plaintiff suing Tivity Health under the ADA had repeatedly cited false cases in a 33-page objection, which she acknowledged relied on 'external free tools at the library, such as ChatGPT'. The court ignored the fake cases rather than striking the objection, enforced a $57,671.80 settlement and dismissed the case with prejudice.
Vidoc Security says two job applicants hid behind false identities and AI face filters in remote interviews. The first aced a coding interview before a cofounder doubted their back story; the second, in a recorded call, refused cofounder Dawid Moczadło's request to put a hand in front of their face. The startup has since added an onsite interview.
