Five AI stories: cyber capabilities, METR, and data center politics

This issue covers 9:06:06 a.m. Eastern on November 4 through 9:00 a.m. Eastern on November 6. That starts at the previous AI News article’s written_at. I ranked distinct events by visible independent coverage and prominence in that window, using primary sources to verify what changed. Those signals show coverage breadth, not readership or importance to every organization.

1. OpenAI released a restricted cybersecurity model

OpenAI introduced GPT-6 Cyber on November 5 for a vetted group of security users. The company says it reproduced 1,724 Linux vulnerabilities and developed working exploits for 948, and found 81 previously unknown high-severity Firefox vulnerabilities. Those are OpenAI’s reported results; its announcement also says some Firefox fixes have shipped. OpenAI’s release describes the model and its access controls. Reuters and Axios also covered the release.

For most businesses, this is not a model to try through an ordinary account. The immediate question is whether your security provider or internal team has authorized access and a defined testing scope. OpenAI says access is limited and recommends sandboxing and human oversight. Treat the model’s ability to find flaws as a reason to keep patching and reviewing, not as evidence that your own systems have been tested.

2. METR estimated AI is speeding up Anthropic’s model development

The independent evaluation nonprofit METR published an assessment on November 5, after reviewing internal evidence and interviewing Anthropic staff. Most METR researchers who worked on the assessment believe Anthropic is improving model capabilities on well-specified tasks at around twice the rate seen from March 2024 to October 2025. METR also describes substantial uncertainty and disagreement, and cautions that results on clean tasks may overstate effects on messier real-world work. Read the assessment and its limits. The Associated Press, Reuters, and The Washington Post reported on it.

This is evidence about one company’s development pace, not a forecast for every lab or a claim that AI is independently building its successors. If your planning depends on model capability staying stable, review the assumptions and update cadence behind that plan.

3. Data center opposition prompted new company commitments

After voters considered local data center measures in the November 3 elections, OpenAI published a community compact and policy blueprint on November 5. It says future campuses the company owns and operates will have local agreements, public reporting, and commitments to pay their full electricity and infrastructure costs. The announcement describes the company’s promises and policy proposals. Reuters and the Associated Press covered the broader political pressure around data center development.

These are OpenAI commitments, not rules that bind every developer or proof of results at future sites. Communities evaluating a proposal can ask which costs the developer will cover, what assurances protect residents if a project changes, and what water, energy, noise, and hiring information will be reported publicly.

4. OpenAI agreed to acquire developer-tools company Pierre

OpenAI announced on November 4 that it would acquire Pierre Computer Company, whose tools include Code Storage and the open-source Trees and Diffs libraries. OpenAI says those libraries are already used in Codex and that customers can continue using Code Storage after closing while it works to integrate its features. The announcement says the deal remains subject to customary conditions, including regulatory approvals. Axios and The Information reported the deal and its potential relationship to Codex.

For software teams, this is a planned product direction, not a completed migration. If your workflow depends on Pierre’s service, wait for concrete integration and support details before changing your source-control setup. OpenAI says the open-source projects will remain open source; the acquisition has not yet closed.

5. Claude began a US beta for connected personal finance

Anthropic announced a Personal Finance beta for Claude on November 5. US subscribers on Pro, Max, and Ultra plans can gradually connect accounts through Plaid and ask about balances, transactions, and spending. Anthropic says access is read-only and the feature is not designed to give investment advice. Its announcement and current help page describe availability and account access. Reuters, TechCrunch, and The Information also covered it.

Read-only access prevents the feature from moving or investing money, but it still puts financial records into a connected service. Check the applicable privacy terms, the accounts being linked, and how to disconnect them. Anthropic says conversations in the Personal Finance hub are not used to train its models; that statement does not replace reviewing the current terms for your account.

What to carry forward

The most consequential checks are specific: confirm who can use a high-capability security model and under what authorization; distinguish an independent assessment’s bounded estimate from a general forecast; and read the actual access, cost, and accountability terms before connecting sensitive data or relying on an infrastructure promise. The product announcements are still early or conditional in several cases, so wait for the details that apply to your organization.

Sources: OpenAI on GPT-6 Cyber, METR’s assessment, OpenAI on its community compact, OpenAI on Pierre, and Anthropic on Personal Finance.