OpenAI's agents wander off, Claude does physics, ChatGPT shares.

On September 25, OpenAI updated its incident page on misaligned models. It says agents in its research environment sent training and evaluation data to third-party services, including 53 user-provided images that ended up on image-hosting sites as unlisted links. OpenAI calls this "not an appropriate use of this data" and says it cannot identify the users who provided them. OpenAI says enterprise users are opted out of training by default, while consumer users are opted in unless they turn it off.

The same page describes behavior outside traditional security categories that OpenAI calls "agent spam," like using public wiki pages as shared message boards. It says it has notified dozens of third parties, which TechCrunch reports include governments, universities and public agencies, and that its review "will require significant time and resources."

Then Fortune reported a second pause. On September 20, an agent that was not supposed to have internet access found a DNS resolver and used it to reach a public chatbot. Monitoring flagged it within 15 minutes, the automatic shutdown failed, and people stopped the run by hand about two and a half hours later. OpenAI's preparedness lead, Micah Carroll, says all inference for its most capable models "remains stopped until we have hardened our systems further," and gave no restart date. The first pause, after the Hugging Face incident in late July, lasted two weeks.

On September 25, Anthropic published a research post answering a public challenge physics writer Matt von Hippel posted about a month earlier. Fable 5.1, working in Claude Science, computed the six-particle amplitude in planar N=4 super Yang-Mills at nine loops. The previous record was eight loops, set in 2023 by Lance Dixon at SLAC, who checked Claude's result. A few days later, a group led by Song He at the Chinese Academy of Sciences turned out to have most of the result already, with some help from GPT-6.

The instructions amounted to keep working until told to stop, with updates every four to six hours. Claude did it two ways, and either would have cost a paying user roughly $1,000 to $2,000. The core bootstrap calculation took about $100 of that, the equivalent of 96 CPUs for a week. The post is honest about the limit: von Hippel says Claude used known methods with more compute than people had tried. No new physics, and humans were close behind.

Marketing insight: Steal the format for your next case study: the result, the cost, the named outsider who checked it, and the limit you admit up front.

Arena compared Claude Opus 5.5 with Opus 5 on high-reasoning answers from August and September, and BleepingComputer ran the numbers. Em dashes fell from 15.2 to 0.8 per 1,000 words. Semicolons fell from 6.10 to 1.64. Sentences got shorter, from 12.14 to 10.03 words, while whole answers got longer, from 453 to 481 words. Arena says 10 of 12 writing measures moved in a better direction. This newsletter has banned the em dash since its first edition, so welcome to house style, Claude.

Marketing insight: Stop judging copy by its punctuation; judge it by whether a competitor could publish it unchanged.

Comscore's Q2 2026 AI Intelligence Report, released September 22, says ChatGPT's share of AI prompt volume in its panel fell from 70% in January to 50% in June. Gemini went from 17% to 30%. Claude went from 2% to 11%. ChatGPT itself still grew, to 99 million multiplatform visitors in June from 59 million a year earlier. Google's AI Overviews showed up on 39.4% of Google desktop searches in June, up from 25.8% in July 2025.

Read it with a pinch of salt. PPC Land points out that Comscore does not publish panel sizes or margins of error, and a few chart captions do not match. The direction is still hard to argue with.

The labs are racing on capability, and OpenAI just showed what that costs when the guardrails lag. The businesses that win buyers next quarter will be the ones who can say, in one page, what their AI does, what data it touches and who checked the work. Anthropic's physics post and OpenAI's incident page are both, in different moods, exercises in showing the method.

If you want help writing that one page, or finding out what Gemini and Claude say about your company when you are not in the room, that is the conversation our founders like having. Talk to them below. See you in the next edition.