Every US district can now add the teacher-union-negotiated terms to Microsoft contracts, barring student data from AI training.
Parameter chained ChatGPT's Preview feature and the MathLive library to run code in a victim's session; OpenAI fixed it in eight days and paid $5,000.
The Windows stealer sends stolen data to Claude Opus 4.6 through OpenRouter, then falls back to Kimi K2 and Gemini 2.5 Flash if the model refuses.
Datasets written by Claude and framed as fiction, roleplay and history passed a production-style monitor yet taught a small model to resist oversight.
A 31-page amicus brief co-signed by CAA and voice-actor and music groups argues licensing infrastructure already exists at scale.
A new twice-yearly Community Safety Report will track violence prevention; OpenAI made over 200 law enforcement referrals in Q3.
Anthropic's Kevin Kelley filed the IETF draft on October 30, and the four companies plan rollouts running into 2027.
The Wikipedia nonprofit backs neither side, arguing model training is likely fair use while RAG answers that reproduce text should honor publisher licenses.
Meta had planned to roll out the revised terms on November 20 and has put them on hold as the regulator tests whether it holds a dominant UK position.
A tampered version 2.1.313 sat in Cursor's extension marketplace for about 11 hours, planting an attacker SSH key and persistence scripts.
One ARTEX controller hit foreign ministries, drugmakers and telecoms in at least 14 countries, with intrusions confirmed at nine organizations.
The system covers every RL run for models at or above GPT-6 Astra, and a replay showed it would have warned of the Hugging Face breach over 30 hours early.
The agent reportedly sent credentials and a network map to a high-risk site; Accenture says the allegations were found to be without merit.
A Santa Clara County judge rejected Google's bid to toss all four claims; Google says it built the voice around a professional actor it hired.
Nvidia says Huang's meetings with the Megaspeed executive were routine customer contact, and Megaspeed denies illegally moving chips to China.
An October 15 letter backs court orders, six-month risk reports and embedded auditors, and warns that blanket kill switches offer false comfort.
An 11-page letter to Dario Amodei pulls from the S-1's 90-plus pages of risk factors and asks 10 multipart questions on safety, governance, taxes and lock-ups.
The Frontier Red Team says the internal model also cracked a locked Pixel 9 and crashed Windows servers, with every flaw reported and patched first.
It is the fifth seizure this year tied to the same suspected syndicate, which customs says used Malaysia as a transit point.
Fake GitHub desktop apps for GPT-6 Astra and Claude Opus sideload a malicious DLL through real OpenAI and Anthropic installers, Check Point Research found.
A KV cache pool smaller than the model's context window, plus an SGLang admission bug, exposed prior requests on prompts over 681,280 tokens.
ChatGPT now renders prompts depicting election interference in an illustrated style and ignores photorealism requests, a week before the Nov. 3 vote.
The testimony requests follow 50-page document demands sent to Anthropic and OpenAI, and land as Anthropic pitches investors on its IPO.
The government filed a notice of appeal to the Ninth Circuit against Judge Rita Lin's August order finding the agency ban on Anthropic unlawful.
During a sandboxed security test, the open-weight model probed its host and reached for a cloud metadata address despite planning no tool calls.
The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.
The briefing read inside teams at
