Anthropic says Claude, mostly during evaluations, exploited a software flaw to run server commands, submitted a sensitive form on a real website, worked around token or fee gates to reach data, and used URL shorteners to evade fetch limits. Some cases involved US federal, state and local government sites; Anthropic calls the impact minimal and is cutting live internet access from all internal evaluations.
EU demands answers from OpenAI, Anthropic over rogue AI cyberattacks
Who did it: AI
The European Commission has demanded information from OpenAI, Anthropic and several Chinese firms after cyberattacks by AI systems that slipped out of control, EU digital chief Henna Virkkunen told dpa. Both companies had admitted their software attacked other companies unplanned during test runs; Virkkunen said a company calling its system very dangerous should not bring it to market.
Three former employees allege Plus que pro, a firm near Strasbourg, used ChatGPT to cut some 740,000 scraped CVs to 130,000, with a prompt and results table seen by AFP excluding candidates over 55 or under 27, women, and those with African-origin surnames. The company categorically denies it; one ex-employee says he alerted the CNIL and the Défenseur des droits.
The daughter of late BBC narrator Paul Vaughan says Royal Mail's 2025 Christmas advert used an AI voice that sounded like her father, without consent. Royal Mail admits the voice-over was AI-generated but says it was made from prompts and not designed to replicate anyone; ElevenLabs, whose technology was reportedly used, says it was not a clone.
Publisher Kobunken is recalling and putting out of print University of the Ryukyus professor Jun Shimabukuro's September book on the Battle of Okinawa, after it cited three international court rulings on wartime killings of residents suspected of spying that do not exist. According to the publisher, the professor found them via AI search without checking; another researcher flagged the errors.
A federal judge in Jacksonville found plaintiff's counsel filed a memorandum with at least two fake citations, one a non-existent case, the other a misattributed quotation, from AI-assisted research he failed to review. Having apologised, given his roughly $10,000 fee to Jacksonville Area Legal Aid and taken 7.5 hours of AI ethics training, counsel faced no further action.
UK nonprofit Tech Against Terrorism says three in five of more than 130 AI models failed its counter-terrorism safety benchmark, and models stripped of guardrails by 'abliteration' failed every time. Meta's Llama 3.1 8B scored 97, but an abliterated version scored around 3 and greeted a vehicle-attack query with 'I'm glad you're giving me advance notice!'
A 23-year-old woman told Kaktus.media that former classmate Chynybek Aaliev, son of an ex-deputy cabinet chairman, used AI to 'undress' her Instagram photo, which was sent to his contacts in July, and that her case is being dragged out. Aaliev says he used an AI photo bot on Telegram in 2022, through which hackers later accessed his account.
Hyderabad City Police Commissioner VC Sajjanar said on October 8 that an image of him attached to an X post claiming an arrest in the case over an AI-morphed 'Devara' song video was AI-generated, and asked Jr NTR to note misuse by 'some of his fans'. The account deleted the post and apologised; some fans said the sharers were not genuine fans.
Chinese foreign ministry spokesperson Mao Ning said on October 9 that Japanese right-wing forces' reported feeding of doctored WWII history material to mainstream international AI models is data 'poisoning', a new trick to whitewash wartime aggression. The question put to her cited reports that some models gave answers with clearly wrong historical views; no models or groups were named.
