OpenAI ships a security model it classifies in its highest risk tier, outside researchers clock Anthropic's pace of progress doubling with AI's help, Claude gets read-only access to your bank account, and Google pledges $17 billion - and its own power bill - for a Wyoming data center. Each explained in plain English, with links to go deeper.
1OpenAI releases GPT-6 Cyber, a security model it rates in its highest risk tier
OpenAI released GPT-6 Cyber on Thursday, a specialized model for authorized security research: finding software vulnerabilities and building working exploits to prove they are real. OpenAI says it scores 48.5% on its ExploitGym benchmark, up from 32.1% for GPT-5.6 Cyber, and that it stayed inside the scope researchers set in 99.93% of test runs without its safety prompt, and 100% with it. The company says the model has already been used to develop 948 privilege-escalation exploits against Linux and to find 81 previously unknown high-severity vulnerabilities in Firefox - many of which Mozilla has now fixed.
In plain English: this is an AI that is very good at breaking into software - on purpose, so defenders can fix the holes first. OpenAI only lets vetted organizations use it, through its Daybreak Red program with identity checks and security requirements, and prices it well above its general models: $20 per million input tokens, $100 per million output.
Why it matters: OpenAI now classifies this model's cyber capability as Critical, the highest tier in its own Preparedness Framework - meaning the company thinks such models are becoming genuinely dangerous in the wrong hands at the same time as they become genuinely useful for defense. The bet behind Daybreak is that the good guys get there first: Cisco, Calif and CodeRabbit are named as early users.
2Outside researchers: AI has doubled Anthropic's pace of progress
METR, an independent AI evaluation nonprofit, published an assessment Thursday after 11 weeks of unusual access inside Anthropic - internal training data, benchmark scores, and staff interviews from August 12 to October 29. Its headline finding: Anthropic's model capabilities on well-specified tasks are now improving at about twice the rate of the previous year and a half, roughly two years of progress compressed into one. METR attributes a majority of the acceleration (its rough guess: 80%) to Anthropic's own use of AI in its research, and says it was not caused by simply scaling up training compute.
In plain English: this is the "AI helping build better AI" question, measured from the inside for the first time. The answer so far: AI is mostly speeding up the human researchers rather than running research by itself - but the assessors saw no obvious bottleneck that would slow the trend over the next year.
Why it matters: faster self-reinforcing progress is the scenario many safety plans assume they will have years to prepare for. The big caveat: Anthropic redacted much of the underlying evidence METR found most informative, so the public is asked to trust conclusions it cannot fully check. METR itself flags wide uncertainty - the true acceleration could be anywhere from 1.3x to 3x.
3Claude can now read your bank account - but not touch it
Anthropic launched a Personal Finance feature in Claude on Thursday: US subscribers on Pro, Max and Ultra plans can connect bank, credit card, loan and investment accounts through Plaid - the same connection layer many budgeting apps use - and ask questions about their real money: where did it go last month, which subscriptions am I paying for, what is left before payday. Access is read-only: Claude can see balances and up to 24 months of transactions but cannot move, manage or invest money, and Anthropic says Personal Finance conversations are not used for training and carry no ads.
In plain English: instead of pasting numbers into a chatbot, you connect your accounts once and ask questions in normal words. Plaid handles the bank login, so Claude never sees your bank password.
Why it matters: chatbots are moving from answering questions about the world to answering questions about you. OpenAI added a similar Plaid connection to ChatGPT in May, The Verge notes. Money is a trust-sensitive place to make that move - read-only access, no training on your chats, and no ads are the table stakes for people to accept it.
4Google plans a $17B Wyoming data center with its own power plant
Google announced its first Wyoming data center on Thursday, a long-term $17 billion investment in Laramie County alongside Tallgrass Energy, Black Hills Energy, AEP and Related Digital. Power comes from a 702-megawatt Bloom Energy fuel-cell plant - the largest development of its kind planned - with carbon capture and storage at Tallgrass's nearby sequestration hub, and the site will be air-cooled to limit water use. Google says it will pay for 100% of the power it uses, cover the infrastructure its growth drives, and generate hundreds of millions of dollars in local property taxes.
In plain English: AI data centers now need so much electricity that building one means building a power plant too. Google is betting on fuel cells, which turn natural gas into electricity chemically rather than by burning it, while burying the carbon underground - and promising locals they will not be stuck with the power bill.
Why it matters: this is the second "we will pay our own way" pledge from an AI giant in two days, after OpenAI's community compact on Wednesday. The backdrop is real resistance: Quartz reports more than 200 Cheyenne residents packed a community meeting about the project, and Google and its partners are forming a Community Council in response. The AI buildout's next constraint is not chips or models - it is permits, power and public opinion.
Checked against original announcements and named press reports on November 6, 2026, in the morning. OpenAI's benchmark figures and risk classification are its own; METR's estimates carry its stated uncertainty.
