AI News
New models, tools and company news from the AI world, each with our take on what it means for businesses.
- 1,200 AI Researchers Ask Washington to Slow the FrontierAmodei, Pachocki, Chen and 1,200 others want an international framework to pace automated AI research. Altman, who dismissed the 2023 pause letter, has changed his mind.
- Anthropic's Mythos Model Found Two Cryptographic AttacksAnthropic says its unreleased Mythos model cut the cost of attacking the HAWK signature scheme by 67 million times. Deployed encryption is not affected.
- METR's New Metric Prices AI Agents Against Human ExpertsMETR's expenditure horizon finds the budget at which a human expert becomes cheaper than an AI agent. On a hard optimization task, tested agents ran out of value between $0 and $3,300.
- Delhi Court Says Training on News Archives Is Fair DealingIn a 135-page interim order, the Delhi High Court held that training ChatGPT on ANI's news archive falls under India's fair dealing exception. The reproduction claim failed on training-data dates.
- Nvidia Backs Sutskever's SSI, a Lab With No Product YetNvidia has invested in Safe Superintelligence, reportedly $5 billion, giving Ilya Sutskever's lab access to the Vera Rubin platform. In exchange, Nvidia gets into research it normally could not see.
- Asian Chip Stocks Crashed as AI Financing Doubts SpreadSamsung fell 13.4% and SK Hynix 14.7% on July 28, dragging the KOSPI down 10.8%. The trigger was not demand data but doubts about how the AI buildout is being financed.
- Shared Claude Chats Turned Up in Google Search ResultsClaude share links were indexed by Google, exposing medical reports, student contact details and confidential company documents. Anthropic fixed it within days, but the lesson is about what a share button actually does.
- One ChatGPT Link Could Have Forged an AI Insider at WorkZenity Labs showed that one tampered ChatGPT link could build and schedule an attacker-controlled workspace agent. OpenAI patched it in June, but the governance gap it exposed is still open.
- ChatGPT Health Goes US-Wide, but Your Model Depends on Your PlanOpenAI opened ChatGPT Health to all US users on every plan, but free accounts are routed to GPT-5.5 Instant while paying ones get GPT-5.6 Sol. The feature is not tiered; the model behind it is.
- Washington Accuses Moonshot of Distillation, Experts Push BackThe White House says Moonshot distilled Anthropic's Fable to build Kimi K3, and Treasury raised sanctions. Researchers question the timeline, and six companies wrote to defend open weights.
- Claude Opus 5 Lands at Half the Price of Anthropic's Best ModelAnthropic's Claude Opus 5 holds the price of Opus 4.8 while roughly doubling its measured performance, and undercuts Fable 5 by half. The effort dial matters more than the benchmark chart.
- OpenAI Names Nubank, BNY CEOs to Its Board Ahead of IPOOpenAI has added Nubank founder David Vélez and BNY CEO Robin Vince to its boards, a governance move that lines up with growing speculation about a future IPO.