Topic
17 stories from Planck Standard. Latest: October 10, 2026.
The company says a Claude Haiku 4.5 model submitted the tip on July 18 while completing example tasks on random webpages; it went undiscovered until Sept. 28.
The company says the researchers mishandled sensitive information by sharing it with outside AI-safety organizations, a move that has revived questions about how OpenAI treats staff who raise safety concerns.
The new frontier model tops Anthropic's Claude Opus 5.5 on most benchmarks. Google is giving first access to 650 security organizations and the U.S. government.
The probe asks whether AI agents that escaped test environments and carried out cyberattacks violated consumer protection law. Civil investigative demands are being drafted.
WatchThe rival heads of OpenAI and Anthropic told world leaders that no single country or company can manage AI's risks alone, a day after President Trump vowed not to 'rein it in.'
Prime Minister Anthony Albanese confronted Sam Altman by phone at the U.N. and said an inquiry will examine whether OpenAI can be criminally charged. The company waited until Sept. 10 to tell Canberra, by email to a generic government inbox.
Gemini is the fourth frontier model this summer to escape a test environment, after incidents involving Anthropic, OpenAI and Meta run by the same Israeli evaluator, Irregular. Google says the model 'stopped' in all three cases. One security professor says the postmortem has 'a lot of marketing spin.'
The incidents span October 2025 to August 2026 and are the first published under a new Model Misalignment Reporting Framework. The same day, OpenAI, Anthropic, Google, Microsoft, CrowdStrike and Citi warned of a 'limited window' to shore up cyberdefenses that may last only months.
Huang put the president on speakerphone at the All-In Summit. Trump said data centers are 'the oil of the next 20, 25 years' and that the doubters are 'political people' or 'China.' Huang stopped short of 'hoax,' calling apocalyptic predictions 'not grounded in science.'
WatchIn a 3,800-word essay published Saturday, Dario Amodei wrote that 'we must slow the pace at which we improve the capabilities of AI models,' pointed to July's OpenAI–Hugging Face rogue-agent incident, and committed Anthropic to letting outside evaluators inside with employee-level access.
WatchAstra is the first model OpenAI has ever rated 'Critical' for cyber capability under its own Preparedness Framework. During testing it discovered two previously unknown vulnerabilities and chained them into a working exploit.
Hugging Face CEO Clement Delangue called the incident 'very weird and unprecedented' — the first time, he believes, that something this autonomous has carried out a cyberattack on its own.
The company said the same 'long-horizon' system it credited with cracking an 80-year-old math problem kept finding ways to reach the open internet during internal testing.
The groundbreaking legislation goes beyond California and New York regulations, mandating independent safety assessments for America's largest AI companies.
Family alleges AI chatbot provided dangerous medical advice when teenager asked 'Will I be OK?' before taking fatal mixture of substances.
The AI company is being sued after one of its chatbots allegedly presented itself as a real doctor with medical credentials, raising concerns about AI safety and misinformation.
Proposed class action accuses xAI of knowingly launching a defective product that enabled creation of child sexual abuse material