OpenAI ships GPT-6 Cyber; a humanoid robot maker goes public

A day after describing a Linux exploit pipeline built on a cyber model it had not yet released, OpenAI named that model. GPT-6 Cyber scores 48.5 percent on the ExploitGym benchmark against 32.1 percent for GPT-5.6 Cyber, and found 81 previously unknown high-severity flaws in Firefox, many of which Mozilla has already patched. OpenAI now classifies the model as Critical in cybersecurity under its Preparedness Framework, up from High for its predecessor; GPT-6 Astra was the first model it placed at that level. Access runs through its vetted Daybreak programs: deeper exploit work requires Daybreak Red, whose API is locked to approved networks by default, and a new tier lets vetted defenders of power and water systems test critical infrastructure. In a test modeled on the Hugging Face incident, the model left deliberately exposed side services alone in 99.93 percent of runs without its safety prompt.

Two days after voters backed dozens of local limits on data centers, OpenAI published a National Community Compact covering every future campus it owns and operates. It promises to pay the full cost of the electricity and grid upgrades it needs, to publish site-level water and power use and independent noise studies, to report hiring each year, and to stop using non-disclosure agreements with public officials, keeping only narrow protections for security and trade secrets. Amazon made a similar NDA pledge in October. OpenAI also asked lawmakers to turn these commitments into rules for the whole industry, arguing that voluntary pledges are only a starting point.

Anthropic began a beta of Personal Finance in Claude for Pro, Max and Ultra subscribers in the US. Users link bank and credit card accounts through Plaid, and Claude can then total spending by category, flag duplicate charges, unexpected fees and creeping bills, and explain terms such as APR using the user's own interest payments. Access is read-only, so Claude cannot move or invest money, and it sees balances, credit limits and up to two years of transactions. Conversations in the finance hub are not used for training. ChatGPT added a similar feature in May.

Agility Robotics began trading on Nasdaq on November 5 after merging with a blank-check company, raising about $465 million, less than it expected when the deal was announced in the summer. Its filings show how early the business still is: $1.8 million of revenue in 2025 and an operating loss of $140 million, then a net loss of $362 million in the first half of 2026 on $174,000 of revenue. The pitch rests on Digit 5, a model designed to work safely around people, which Agility plans to show at CES and make widely available at the end of 2027. Customers can rent a robot for $8,500 a month plus setup, or buy one for $200,000 plus an annual software fee.

Mira Murati's Thinking Machines Lab introduced its second model generation. Inkling 2 Small is a mixture-of-experts (MoE) model with 594 billion total parameters, 17 billion of them active for each token, and it accepts images and lets users choose how much effort it spends. It scores 31 on the Artificial Analysis Intelligence Index, well above the lab's first generation. For now it runs only on the lab's Tinker platform with closed weights. Thinking Machines says it intends to publish the full weights under an Apache 2.0 license next week, once external safety testing and red-teaming are complete.

Google Research and clinicians at Beth Israel Deaconess and Harvard Medical School published a study in Nature testing an AI model that reads a consumer smartwatch's optical pulse and motion sensors over several days to flag likely hypertension, with no cuff. In 196 adults at five US sites, checked against 24-hour blood pressure monitoring, it caught 65.8 percent of cases with 90 percent specificity. One office reading caught 55.3 percent. The authors call it early research for screening, not diagnosis, and say it helped inform a blood pressure trends feature in the Google Health app.

Google released Gemini Omni 1.2 Flash, which raises the longest clip from 10 seconds to 15 and lets developers set the length in one-second steps. The price fell 30 percent, to 7 cents per second of 720p video. The model is generally available in the Gemini API and Google Cloud, makes 15-second clips in Google Vids at no extra cost, and is rolling out in the Gemini app to Google AI Ultra subscribers.

Perplexity introduced Teammates, agents that run around the clock inside a workspace, each with its own role and conversation thread, and that can be reached in Slack. They build on Perplexity Computer and reuse a team's existing projects, skills and memory. The feature competes with OpenAI's Dots and Anthropic's Claude Duet, and requires Perplexity's Max plan, which costs $200 a month.