Daily Brief Oct 23, 2026
How we checked the items below
For readers who want to dig in. What happened, in brief, is on the Today tab.
This issue draws on last night's sweep. Main-line item 1 rests mainly on a summary from METR, an outside evaluator that labs commission for risk assessments; that item's judgment paragraph also cites METR's October 20 review, so both documents come from the same organization. Main-line items 2 and 3 are both the parties' own announcements, with no third-party reaction. Most events in the main line and the later sections are dated October 22; the research preprint is dated October 19, and From the archive revisits an older topic. Last night's sweep took in 608 pieces, and this issue uses 7 outside sources with checkable links.
1. [Today] (announcement, October 22) METR estimates 75% to 94% of OpenAI's internal agent traffic from GPT-5.5-class models was in scope for monitoring; reinforcement-learning training was not
Not proven yet： the readings cover the system before July 21; current coverage wasn't measured
How we checked
METR's original was read, and the quotes were checked word for word. This is METR's own summary. The public version is a short one agreed with OpenAI in advance. OpenAI alone made the cuts, and METR had no say over what was removed. The 75% to 94% is the in-scope share that METR's two staff members estimated from the access they had. In scope means the traffic was meant to be sent to monitoring. It doesn't mean all of it completed monitoring. METR writes that the uncertainty is large in both directions. METR also writes that system overload, monitor errors and software bugs could all keep in-scope traffic from completing monitoring. The share excludes agent traffic on OpenAI's internal research inference service. The readings cover only the system before July 21. OpenAI expanded monitoring afterward, and the exercise didn't measure coverage after that expansion. Setting a header and starting a new agent could themselves be caught by monitoring. METR describes most bypass routes as possible in principle. METR had no access to check whether any were actually used. The monitoring-request error rate below 0.2% is data OpenAI gave METR. The three OpenAI posts METR links to, including "Hugging Face incident and the road ahead," haven't been reviewed and may describe the expanded monitoring scope. METR also ran a monitoring stress test with Anthropic in March; that original hasn't been reviewed.
2. [Today] (announcement, October 22) An Anthropic-run version of Claude on Google Cloud reaches general availability at the same price and on the same bill, with Anthropic as data processor
Not proven yet： Google's post doesn't say whether the Anthropic-run version has data residency
How we checked
The full Google Cloud announcement was read, and the quote was checked word for word. This is Google's own announcement; Anthropic's documentation page for its side hasn't been reviewed. "Same pricing" comes with no price list. The post doesn't say how much later new features reach the Google-run version. The post lists data residency and customer-managed keys only under the Google-run version; whether the Anthropic-run version has them depends on Anthropic's own documentation.
3. [Today] (announcement, October 22) OpenAI says Uber is opening ChatGPT Work to more than 30,000 employees after a 1,000-person pilot, with no results published
Not proven yet： 30,000+ is the number of employees given access, not active users
How we checked
OpenAI's full announcement was read, and the quote matches it word for word. This is a vendor's customer announcement in which Uber is the party quoted; Uber's own announcement hasn't been reviewed. Calling this Uber's first company-wide deployment of an agent platform for knowledge work is OpenAI's wording. The post doesn't say whether Uber also uses other vendors' agent platforms. The post contains no cost, usage or productivity figures. The Uber partnership page and the GPT-6.1 Sol launch page that OpenAI links to haven't been reviewed.
[Today] (research post October 22) Anthropic writes that Claude proved 239 fuzz-testing entry points in 123 open-source libraries can't crash on any input up to 64 MB. The proofs cover single-threaded static RISC-V executables only.
How we checked
All the figures are Anthropic's own, and Anthropic has published some of the proofs for independent checking.
[Today] (announcement October 22) Anthropic donates Agent Skills, its format for agent skills, to the Agentic AI Foundation under the Linux Foundation, and says more than 70 agent platforms have integrated it.
How we checked
Anthropic also writes that 1 million skills have been published and installed nearly 280 million times; the post doesn't say how these are counted. The integration and install figures are Anthropic's own, with no second-party check. The post gives no data on how well each vendor's agents use skills.
[Looking back] (SecondSource deep dive, October 15) The CPU racks for running AI agents shown at the October 12 Open Compute summit came from four companies, AMD, NVIDIA, Intel and Arm, and those products were announced before the summit.
How we checked
Still missing today. The gap that would most affect judgment is OpenAI's current monitoring coverage after the expansion: METR's exercise didn't measure it, and the three OpenAI posts METR links to haven't been reviewed. Other gaps: the original of METR's March monitoring stress test with Anthropic, Anthropic's documentation page for the Anthropic-run Claude, Uber's own announcement, the full Bashi AI paper and the Agent Skills specification page haven't been reviewed. The appendix and expert commentary on Anthropic's formal-verification post, and new product posts published yesterday by Databricks, NVIDIA, Cohere and OpenAI, haven't been read yet.
Sources & accounting
The past 24 hours. 608 new pieces came in from October 22 to 23: 346 academic papers, 131 social-platform posts, 60 company or personal blog posts, 55 from a second academic-paper feed, 8 subscription newsletters, 5 company filings and 3 show transcripts. Of those 608, 85 were read in full and 56 were screened out as off-topic; 467 haven't been read yet. Of the 346 academic papers, 10 were read and 333 haven't been read; of the 60 blog posts, 4 were read. None of the 5 company filings has been read: quarterly reports from IBM, Lam Research and Meta, and one foreign-issuer filing each from Nebius and TSMC. Accounts with the most posts included @teortaxesTex, @bhorowitz, @TheStalwart, @saranormous and @geoffreyirving. The 633 original X posts pulled last night are a different population from the 131 new posts above, so the two figures can't be divided. Shows included Latent Space, Training Data and Dwarkesh; newsletters included Latent Space, Gary Marcus, Dwarkesh, Newcomer and The Zvi. A full set of older CoreWeave blog posts reappeared in the feed at the same moment; that is a replay at the source, not a new event. Data Center Dynamics and the Microsoft Research Blog weren't fetched last night.
One-time catch-up. No new one-time catch-up of older material today.
Tracked sources. The long-term roster currently has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters, with the rest spread across earnings reports, keynotes and other channels. Note: those are counts of tracked sources, a different population from the 608 new pieces above; they can't be added or compared. Representative names: on social platforms, Ben Horowitz, Miles Brundage, Gary Marcus, Elon Musk and Kevin Weil; among newsletters, Zvi Mowshowitz, Dwarkesh Patel and Eric Newcomer; on the shows roster, interviewees such as Sam Altman, Dario Amodei and Demis Hassabis. This issue uses 7 outside sources in the body. Only outside links actually cited in the body are counted, and links to SecondSource's own site are left out; this is a different population from the 529 tracked sources above. On the Taiwanese side, 7 new sources were added in September, bringing the total to 134; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.