Good morning. It's Friday, November 6th.

OpenAI shipped its first model rated Critical for cyber

OpenAI released GPT-6 Cyber on Thursday and classified it Critical under its Preparedness Framework, the top cyber tier, one notch above where GPT-5.6 Cyber sat. On OpenAI's own ExploitGym evaluation it scored 48.5% against 32.1% for its predecessor. Access is restricted to vetted Daybreak Red customers through the API and Codex, with a wider Red rollout promised over the coming weeks.

OpenAI says the model surfaced 81 previously unknown high-severity vulnerabilities in Firefox, and The Register reports Mozilla has started shipping fixes. The companion writeup, A thousand fewer paths to root, describes an October campaign that analyzed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 of them inside 30 hours, and ended with 948 working local privilege escalation exploits, 588 of them previously unknown. Total bill: $1,213,430 in API credits, about $1,279 per exploit.

On scope discipline, OpenAI says the model stayed inside its authorization boundary over 99% of the time and that 0.07% of cases went out of scope. It declined to say what happened in those. Worth remembering that GPT-6.1 Astra was shelved in September for overstepping authorized boundaries, which is why the number is in the release notes at all.

Read more.

Data centers lost 40 of 41 ballot measures

Tuesday's midterms put data centers directly on the ballot 41 times across ten states, and voters backed a restriction or an outright ban in all but one. Ohio carried 21 of them: Conneaut, Defiance, Oregon and Sunbury capped data centers at 25 megawatts, while Granville, Hubbard, Urbana, Upper Sandusky and Piqua banned them outright. All seven Illinois measures went against development. A Wisconsin measure now requires a public vote on any development over $450 million. A Bowling Green State University poll last month had 71% of Ohioans backing a temporary construction ban, so this was not a surprise to anyone who was looking.

Markets caught up on Thursday. The Nasdaq fell 5.52% to 24,315, its worst day since April 2025 and into correction territory, more than 10% off its October 6 record. The S&P 500 dropped 4.36% to 7,353.73 and the Philadelphia Semiconductor Index lost more than 12%. Nvidia closed down 9.26% at $181.60, AMD fell 17.26%, Broadcom 10.43%, and the neocloud names IREN, Applied Digital and Nebius each shed more than 14%. eToro's Jakub Rochlitz read it simply: the market now expects the buildout to be "slower and more expensive."

Read more.

OpenAI and Meta each bought a coding startup on the same day

OpenAI agreed on Wednesday to acquire Pierre Computer Company, which builds a Code Storage service plus the open-source Trees and Diffs libraries that Codex already uses for file tree navigation and visual diffing. Terms were not disclosed and the deal is subject to regulatory approval. The plan is to pull Code Storage into Codex, turning it from a completion tool into a repository and code review environment. Axios framed it as a run at GitHub, which tracks with Pando, the Git-compatible backend we covered Wednesday. OpenAI says the open-source projects stay available.

Hours later Meta acquired Conductor, founded in 2024, which runs multiple coding agents in parallel workspaces across cloud, local machines and phones using models from Anthropic, OpenAI and others. Terms undisclosed. Founder Charlie Holtz said he is looking forward to working with Alexandr Wang and Nat Friedman at Meta Superintelligence Labs, and teased something new in early 2027. If you use it: the cloud service shuts down November 20 and user cloud data gets deleted after that, with core features running until early next year.

Cat Wu, who led Claude Code product and left Anthropic on October 30, has joined OpenAI to run product management for Core Products, under a week after walking out the door.

Read more.

The most useful shipments this week are unglamorous. Claude Code background agents now write their status updates with Haiku 5.5 instead of the session model, which Anthropic says cuts the cost per update about 99% and trims roughly 5% off session allowance use. If you run Claude in a company that has a security team, the bigger one is Anthropic's new prompt and pre-tool-call hooks, now generally available: prompts and proposed tool calls get routed to your existing DLP vendor before they execute, so "we cannot let Claude touch that data" stops being a reason to not deploy. A security center and a cross-org parent console are in beta alongside it.

On the consumer and agent side: ChatGPT Images 2.6 generates animated GIFs from prompts or photos and keeps reference images persistent, with Flare and Sunburst variants in the API. Notion opened shared memory so ChatGPT, Claude and other assistants read the same workspace context, which is the first credible answer to agent context living in five places. ElevenAgents added native connectors for NiCE CXone, Genesys Cloud, Amazon Connect and Five9. Pangram Image 1.0 detects AI-generated images with heatmaps showing where it is confident, which is more than most detectors will tell you. And xAI demoed Grok Bot canceling a subscription over the phone after user approval, which is either the best or worst product demo of the week depending on your relationship with your internet provider.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.