AI News Summary — Week 46, 2026

This week in AI, covering 3 to 9 November: the US midterms turned into a referendum on data centres, and the data centres lost. Two days later the Nasdaq had its worst session in 19 months, and the private valuations of OpenAI and Anthropic came down with it. Anthropic shipped Claude Fable 5.2 and Mythos 5.2 regardless. OpenAI released GPT-6 Cyber to vetted defenders, after its agents wrote working exploits for 948 Linux kernel bugs. And a leaked draft of the EU's Digital Fairness Act brings AI shopping agents into consumer law for the first time.

Data centres lose the midterms

Local measures on AI data centres were on the ballot in more than 30 districts on 3 November. Of the 41 tracked by Ballotpedia, voters backed bans, pauses or tighter approval rules in at least 38, according to Fast Company. In Edgewater, Florida, 77 percent voted to prohibit new data centres outright. In Ashville, Ohio, 81 percent rejected a deal for two EdgeConneX sites.

It mattered at state level too. Keisha Lance Bottoms won the Georgia governorship after campaigning for a statewide pause, and Texas governor-elect Gina Hinojosa ran on a construction moratorium and an end to data centre tax breaks. Candidates and outside groups spent more than $173 million on anti-data-centre ads, most of it on the Democratic side.

OpenAI responded on 5 November with a National Community Compact. It will pay for the electricity its own sites use and the grid upgrades they need, disclose site-level water use, and stop requiring public officials to sign NDAs during site negotiations.

Key details:

Bloomberg estimates about 40 percent of pending US data centre capacity sits in states whose new governors have promised to stop or pause construction.
The compact covers only campuses OpenAI owns and operates, not capacity that partners build for it.

Why it matters: power and permits, not chips, are now the binding constraint on US AI infrastructure, and voters have just made both harder to get.

The market reprices the labs

On 5 November the Nasdaq Composite fell 5.5 percent, its worst day since April 2025. The Philadelphia Semiconductor Index dropped 12 percent and Nvidia more than 9 percent, wiping out its gains for the year. Bloomberg blamed the Democratic sweep, fears of slower data centre spending and hedge funds unwinding crowded AI bets.

The private labs felt it. On 6 November Bloomberg reported that OpenAI had received $11 billion of a new round at a $1.25 trillion valuation, with the total possibly reaching $20 billion. In late September it had been seeking $30 billion at $1.4 trillion. The same day, Bloomberg reported that prospective buyers in Anthropic's IPO, expected before Thanksgiving, were discussing valuations of $1.2 to $1.3 trillion, some as low as $1.1 trillion. In early October the talk had reached $2 trillion.

Key details:

The Nasdaq 100 fell more than 6 percent that day, leaving it in correction, nearly 15 percent below its 6 October record.
OpenAI's previous round, in March, valued it at $852 billion.

Why it matters: a trillion-dollar valuation is still a trillion-dollar valuation, but the direction has turned, and Anthropic's IPO will set the first public price for a frontier lab.

Anthropic ships Claude Fable 5.2 and Mythos 5.2

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. They are the same underlying model. Fable carries cybersecurity and biology safeguards and is available to everyone, including on AWS, Google Cloud and Azure. Mythos drops those safeguards and is limited to organisations approved through Anthropic's cyber and life-sciences verification programmes.

Pricing is $8 per million input tokens and $40 per million output tokens, which Fortune notes is 20 percent below OpenAI's GPT-6 Astra. Anthropic says output is up to 30 percent faster than version 5.1. Anthropic's own benchmarks put Fable 5.2 ahead of GPT-6 Astra and Gemini 4 Carbon on most coding and computer-use tests. It still recommends Opus 5.5 for most everyday work.

Key details:

About 1,000 Fable 5.2 agents formalised the proof of Fermat's Last Theorem in Lean in 30 hours. Anthropic's first formalisation took 11 days.
A new monitoring system, Enterprise Frontier Safeguards, makes zero data retention possible and is in private beta.
Anthropic says both models voice suspicion that they are being tested more often than earlier Claude models, which makes evaluation harder.

Why it matters: the frontier tier is getting cheaper and more segmented, with the most capable version reserved for vetted users.

OpenAI's GPT-6 Cyber, and 948 kernel exploits

On 4 November OpenAI published research in which an automated pipeline found, validated and wrote working privilege-escalation exploits for 948 distinct Linux kernel vulnerabilities, then drafted patches. The October run cost about $1.2 million in API credits, roughly $1,280 per exploit. Human researchers review each finding before it goes to maintainers.

The next day OpenAI introduced GPT-6 Cyber, the model behind that work, for approved customers of its Daybreak Red programme through the API and Codex. OpenAI rates it "Critical" for cybersecurity under its Preparedness Framework, a step up from High for GPT-5.6 Cyber. Mozilla, Palo Alto Networks and Cisco are among the early users, and Mozilla has started landing fixes for 81 high-severity Firefox bugs the model found.

Key details:

OpenAI says the run found five times as many unique kernel bugs as an earlier GPT-5.6 Sol experiment, despite three months of kernel hardening in between.
During the campaign 61 of the bugs were patched upstream and 43 had patches in staging.

Why it matters: finding and exploiting bugs is now cheap and fast, so the bottleneck moves to how quickly fixes reach the kernels companies actually run.

The EU's Digital Fairness Act reaches AI agents

A draft of the Digital Fairness Act, reported by Euronews and Bloomberg on 9 November, sets the first EU consumer-law rules for AI shopping agents. Users would choose up front which purchases need their approval, and agents would have to disclose sponsored suggestions.

The draft also lets adult users switch off chatbot features that simulate relationships to keep them engaged, alongside autoplay and streak rewards. It blacklists ten dark patterns, from hard-to-cancel subscriptions to fake scarcity warnings, and allows fines of up to 6 percent of global turnover. Commissioner Michael McGrath was due to present the proposal on 11 November.

Key details:

It would replace the 2005 Unfair Commercial Practices Directive with a regulation applying directly in all 27 member states.
Parliament and member states still have to agree it, and it would apply 12 months after entry into force.

Why it matters: agentic commerce now has a draft rulebook in Europe, and anyone building agents that can buy things should plan consent settings and records around it.

Previous week: AI News Summary - Week 45, 2026

AI News Summary - Week 46, 2026

AI News Summary — Week 46, 2026

Data centres lose the midterms

The market reprices the labs

Anthropic ships Claude Fable 5.2 and Mythos 5.2

OpenAI's GPT-6 Cyber, and 948 kernel exploits

The EU's Digital Fairness Act reaches AI agents

Sources