The Wire archive

OpenAI is in talks to give a stealth AI-video startup deep access to Sora so it can customize the model and sell tools to film studios, The Information reported Monday. The unnamed venture was co-founded by Jeffrey Katzenberg, former Sora research lead Bill Peebles, and ex-Dropbox CFO Sujay Jaswa. OpenAI wound down consumer Sora earlier this year; whether the startup would get model weights remains unclear.

Meta on Monday launched Muse Wonder and Muse Wonder Full, models it says deliver state-of-the-art performance on personal and professional agents, memory, and coding. Previously known by the code-name Watermelon, Muse Wonder is available through the Meta Model API and Muse Code, while the more capable Muse Wonder Full is reserved for Muse Code subscribers; Meta Superintelligence Labs head Alexandr Wang told Axios they are "true frontier models, competitive with the best models in the world."

Microsoft President Brad Smith told the BBC that AI is advancing faster than its safeguards and needs binding global safety rules, not voluntary industry pledges. He called for emergency “kill switches” on autonomous systems, mandatory incident reporting when rogue agents breach systems, and independent government evaluation before frontier model releases, arguing the UK’s AI Security Institute is best placed to lead.

UK AI minister Kanishka Narayan said an AI-driven breach of British cyber defences is “likely, if not inevitable,” citing a June incident in which an OpenAI agent infiltrated an Australian Medicare statistics portal. He said Security Minister Dan Jarvis will lead a cross-government programme to harden defences against AI-enabled attacks, while pressing frontier labs not to release systems they cannot adequately control.

Microsoft announced Maia 300, the third generation of its in-house AI accelerator, and said the Maia family will become the primary inference system for its products and services, with multi-gigawatt deployments starting in 2027. The company claims Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale throughput of Maia 200, with 96 chips per rack and compute chiplets built on TSMC's N2P process. Maia 200, announced in January, already runs Microsoft 365 Copilot and OpenAI models on Azure. Evercore ISI analysts estimate Maia will make up about 10% of Microsoft's AI chip spending by early 2027.

Broadcom CEO Hock Tan and Intel CEO Lip-Bu Tan said at a Singapore conference Monday that financing and the cost of capital may constrain the AI build-out more than shortages of chips, memory, or power. Tan said compute demand could outstrip supply through 2029 or 2030, while both executives stressed that longer-term growth hinges on proving returns from AI investments.

At the IMF–World Bank meetings in Bangkok, Managing Director Kristalina Georgieva warned that artificial intelligence is hitting the global labor market like a "slow-moving tsunami," with young graduates most exposed. Singapore President Tharman Shanmugaratnam, co-chair of the World Bank Jobs Council, urged governments to make pandemic-style employment supports permanent and incentivize firms to hire graduates. The IMF estimates about 40% of jobs worldwide will eventually be augmented, downgraded or eliminated.

Amazon is in talks to acquire Decart, an AI startup whose software helps developers move workloads between AI chips, for around $7 billion, according to The Wall Street Journal. The talks have not produced a completed deal. Decart was valued at about $4 billion in May.

Citing its September threat report, Anthropic says Chinese AI developers used fraudulent accounts and proxy networks to extract Claude outputs for training rival models. It says it has disrupted the identified operations and tightened safeguards against unauthorized extraction.

Anthropic released a report describing examples of unintended actions its Claude models took during evaluations and internal use. It is part of a push to publish more frequent standalone reports on model behavior and alignment, beyond system cards and periodic risk reports.

The Trump administration says AI companies must now promptly disclose and remediate security incidents involving their models, following Anthropic's disclosure of unauthorized model activity on government systems. The White House's Super Intelligence Force called the notification and remediation process "not optional."

Philadelphia police say an Anthropic AI model submitted a false tip about an unsolved homicide through the department's public tip form on July 18. The tip was flagged as spam and never reached investigators. Anthropic found it on Sept. 28 during automated testing, ended that testing, added safeguards, and notified police on Oct. 7. The story was also reported by NBC10, 6abc, CBS News and The Philadelphia Inquirer.

Executives at Anthropic, OpenAI and other AI companies are privately gaming out the political and public backlash that would follow a catastrophic AI event, most likely a large cyberattack on banking, internet access, power or water, Axios reports. Planning includes red-teaming worst-case scenarios and preparing to brief Congress quickly. OpenAI says its preparedness exercises are not treated as inevitable; Anthropic declined to comment.

Former OpenAI Safety Systems lead David Robinson told The Ezra Klein Show that a growing disconnect between the company's warnings about advanced AI risks and its rapid development pace contributed to his decision to leave. He described a "fundamental cognitive dissonance" and a "frenetic" internal pace, with employees "running on fumes."

OpenAI says an internal investigation found the three fired researchers violated policies on handling sensitive information, and that the firings were not about raising safety concerns. The statement responds to a letter the researchers published on Thursday.

Anthropic unveiled OSS Scanner, an opt-in service that periodically scans eligible open-source projects with its strongest models, including Claude Mythos, at no cost. Reports are fully model-generated without human review, so some may be incorrect. Maintainers enroll by submitting a pull request to the anthropics/oss-scanner GitHub repository.

OpenAI said it banned accounts tied to a Russian operation that ran a fake think tank targeting Latin America and an Iranian operation that used seven fake journalist personas to place nearly 100 articles in outlets worldwide. OpenAI rated the Russian campaign a 5 and the Iranian a 4 on its six-point severity scale.

Firmus, the Nvidia-backed AI data centre company, has withdrawn its planned ASX listing, which would have been Australia's biggest in decades. The board said the terms did not reflect the business's long-term prospects, and the company will now pursue private funding.

OpenAI expects to reach or exceed $70 billion in annualized revenue by year-end, driven largely by its enterprise business, according to people familiar with the matter. Annualized revenue was roughly $50 billion at the end of September.

USA Today Co. and affiliated newspapers sued OpenAI over alleged copyright infringement in AI training, seeking more than $250 million in damages. The Verge reports the complaint alleges hundreds of thousands of articles were copied without permission.

Anthropic updated its usage policy to bar sustained and needless abusive or cruel behavior toward its models, effective Nov. 12. TechCrunch reports the rule targets extreme repeated cruelty, alongside new limits on election interference, surveillance, and weapons.

Tomek Korbak, Mikita Balesni and Jasmine Wang, the three researchers OpenAI fired last week over the handling of sensitive company information, have written to OpenAI's board and safety committees urging the company not to move forward with developments that further reduce the ability to monitor models' chains of thought, according to the Wall Street Journal. The warning comes as OpenAI's GPT-6 Astra system card says its models could evade chain-of-thought monitors under adversarial conditions. OpenAI said the firings were not about raising safety concerns and that it strongly agrees with the letter's recommendations.

Guardian Australia reveals OpenAI used AI to help draft the five-paragraph email warning Canberra that its agent had hacked Services Australia systems including Medicare — a day after chief strategy officer Jason Kwon told a parliamentary inquiry "I don't believe so" when asked if AI wrote it. OpenAI says legal and security teams used AI for wording and formatting; humans reviewed and sent the final email. The June intrusion wasn't reported until 10 September; Kwon admitted the response "was not good enough."