Policy officials from OpenAI, Anthropic, Google and Meta testified by video to the New York City Council, which is considering 10 bills that include outside validation of AI systems and a human ability to shut models down, per CBS New York. Asked by Speaker Julie Menin to estimate the odds of a worst case catastrophe, OpenAI's Morgan Dwyer said the exact figure did not matter because no level between 1% and 20% would be acceptable. Menin called the answer "flippant at best." Former Anthropic engineer Jacob Coxon told members that on the current path it is "more likely than not that humanity loses control to AI," which he said could end in human extinction, and he accused the companies of acting recklessly, per The Wall Street Journal. Another proposal would give whistleblowers a share of fines, AP reported. SpaceX skipped the hearing despite a subpoena, and the city is now pursuing legal action.
Read at AP News ↗ • Read at CBS New York ↗ • Read at WSJ ↗
President Lee Jae Myung told a cabinet meeting that AI seems to have played a role in a string of cyberattacks on South Korean banks and told officials to focus staff and money on limiting the harm, per Reuters. He said the suspected AI use had caused considerable public concern. The attacks exposed customer data at seven financial firms, including Shinhan Bank and KB Kookmin Bank, and police formed a team of 28 investigators, Yonhap reported. Lee Eog-weon, chairman of the Financial Services Commission, the top financial regulator, held an emergency meeting Sunday and ordered firms to inspect every system exposed to the internet. A server used in the attacks carried a Chinese language string linked to ARTEX AI, an open source tool that uses AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. to automate break-in testing. Authorities have not confirmed its use.
Read at Reuters ↗ • Read at Yonhap News Agency ↗ • Read at BleepingComputer ↗
OpenAI Chief Strategy Officer Jason Kwon apologized in person over breaches involving OpenAI's AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. to the Australian Parliament's Joint Select Committee on Artificial Intelligence in Sydney, telling members, "We are sorry, and we know we have work to do to rebuild trust with the Australian people," per The Guardian. Kwon said OpenAI is combing agent training logs back to November 2025 and would tell agencies "very, very quickly" about any further incidents. He conceded the company should have alerted the government sooner and said staff now get warnings when models use the internet improperly during training, ABC News reported. Anthropic safeguards head Dave Orr told the committee the company had searched hundreds of millions of model transcripts and found no unauthorized contact with Australian government systems, though it cannot see the activity of customers whose data it does not keep.
Read at The Guardian ↗ • Read at ABC News ↗ • Read at Bloomberg ↗
Over the coming weeks OpenAI will embed an invisible watermark in text that ChatGPT and Codex produce for eligible users in the European Union, to comply with AI Act transparency rules in force since Aug. 2, per TechCrunch. The method, called textGrain, nudges the model's word choices into a statistical pattern that a detector can read, and it does not identify the user. OpenAI said watermarking will not be a global default, though API customers anywhere can now opt in for select models. In company tests, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI said textGrain matched or beat Google DeepMind's SynthID watermark in its own evaluations. Approved researchers will get detector access first.
Read at TechCrunch ↗ • Read at The Verge ↗ • Read at OpenAI ↗
Insurers are preparing for multimillion dollar claims from AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. acting outside their developers' control, and some industry figures say the heads of OpenAI and Anthropic, Sam Altman and Dario Amodei, could face personal lawsuits, per the Financial Times. Broker Aon reviewed more than 300 legal cases involving AI and found exposure across crime, intellectual property, media, cyber and technology errors and omissions policies, the last of which cover professional mistakes. "Ultimately, the CEO of OpenAI is responsible for the Hugging Face breach because there was a lack of control at the corporate level," said Tim Rayner, UK head of underwriting and claims at Verisk, pointing to directors and officers coverage, which pays when executives are sued over their decisions. No such claim against an AI executive has been tested in court.
Read at Financial Times ↗
New York magazine and Asterisk magazine gathered eight former employees of OpenAI, Anthropic and Google DeepMind, including Jacob Coxon, Daniel Kokotajlo and Miles Brundage, to explain why they left. Coxon said longtime Anthropic research staff had consistently told him "2027 is when things get crazy" and did not expect government regulation to arrive in time. Former DeepMind alignmentAlignmentThe problem of making an AI system reliably pursue the goals its developers and users intend, and the research field devoted to it. Misalignment covers everything from a chatbot flattering users to a capable system deceiving or resisting its operators. researcher Alex Turner, who worked on making AI systems follow their designers' intent, said he urged Google chief scientist Jeff Dean to back alternative contract terms for a Pentagon AI deal. Google signed the deal anyway. "I really don't think companies can be trusted to regulate themselves," Kokotajlo said.
Read at New York Magazine ↗