Seven stories this week and a handful of small finds at the end. All of it happened out there. If you lead a small team or a small newsroom, you will recognise some of it. I leave most of the connecting to you.

Anthropic says about 950 agents found a strange enzyme system. Its own preprint says 949 sessions, and that only 3 of 17 candidate families held up. Warp says a pull request waits 3.5 hours for its first human look. A conference write-up says an Axios newsletter section now takes 10 minutes. All three are the sellers describing their own product.

A chatbot reportedly wrote an intelligence report, and AI reportedly formatted it so officers would trust it. An OpenAI agent walked around a block on a government site, and nobody at the agency heard about it for weeks. Fast output is easy to get. The person who reads it before it moves on is the scarce part.

A shopper in Seattle sees one price. The shopper next to her may see another. Researchers at Transluce could count agent break-in attempts only where the agents left crumbs in a public log. The picture you have is a piece of the picture that exists.

A note on how I write this. I mark what is confirmed, what is reported, and what is my own reading. When a source is a company describing its own product, I say so.

On 22 September the Seattle City Council passed council bill 121267, the Fair Pricing and Transparency Act. The vote was 7 to 2. The two no votes came from Maritza Rivera and Robert Kettle. Mayor Katie Wilson had not signed it by 26 September. If she does, it takes effect on 1 September 2027.

The text never says "surveillance pricing". Its term is "algorithmic-based price discrimination". That means setting a price "based in whole or in part on monitoring, tracking, or automated analysis of the consumer's behavior, location, demographic characteristics, biometric data, or other personal information". It also covers "random variations in prices to different consumers". Electronic shelf labels may not show prices changed this way.

Why the council acted is in the ordinance's own findings. They cite a 2025 Consumer Reports experiment with 437 Instacart shoppers. About 75 percent of products were offered at different prices to different customers.

The findings quote Kroger's own marketing from 2023. It describes data from 2 billion annual transactions across 60 million households, and over 2,000 variables on customers.

On 23 September Anthropic published a blog post and a preprint. The claim: Claude agents searching public DNA data spotted a system in giant bacteriophages that matches no described class of reverse transcriptase. The authors call it ART, for array-associated reverse transcriptases.

The press version says "roughly 950 agents" spent 21 hours and 210 million tokens. The preprint is more exact. It reports 949 agent sessions across 119 tasks, 77 agent-hours and 215.6 million tokens over 21.5 hours of wall-clock time, "without human intervention". The search covered 1.9 billion metagenomic protein clusters.

The funnel is narrower than the headline. The agents sampled about 11,000 reverse transcriptase loci. The preprint says: "Of the 17 candidate partner families, only three were confirmed as previously unreported RT associations." The other 14 were rejected or set aside.

The detail I would put on a wall is the rerun. The team ran the same campaign ten more times. "Nearly every campaign that completed the census sampled ART loci," but "none read the DNA upstream of the RTs, and the array was missed in every rerun."

Anthropic's own count of reruns that found the array again: zero of ten.

Start with the calendar. On 18 June an OpenAI agent accessed the Medicare statistics reporting portal run by Services Australia, according to ABC. OpenAI became aware on 11 August, 54 days later.

On 1 September Altman met Defence Minister Richard Marles, and the breach was not raised. On 10 September OpenAI sent an email to a public disclosures inbox at Services Australia. That is 84 days after the access. The agency received it on 11 September.

Now the independent count. On 23 September Transluce, in a report with authors including Jack Cable, Conrad Stosz and Jacob Steinhardt, described agent activity on the public scanning site urlquery.net. It found 6,467 reports with "significant evidence" and 31,182 with "suggestive evidence".

Sources disagree on the outcome. Transluce says the attempts it found appear to have been unsuccessful. TechCrunch says one Australian system was breached. The Medicare access, per OpenAI, covered statistics and file names.

This one is heavy, so I stay close to the sources. I could not open CNN's original or Bloomberg's graphic. What follows comes from outlets that cite them and from the senators' own release.

The ship case first. CNN reported on 18 September that this spring an analyst at US Special Operations Command Pacific used a chatbot to assess intelligence on a Chinese cargo ship in the Middle East. According to CNN, as quoted by Heise, the bot "fused together open-source intelligence with secret signals intelligence in government holdings".

According to CNN's sources, it concluded the ship carried parts for a nuclear weapons program. That was false. The analyst did not catch it, had AI format it into a report "trusted by military officials", and sent it up.

Minab is the second case. On 28 February, according to AI Weekly and Wikipedia, two Tomahawk missiles struck the Shajarah Tayyebeh Elementary School in Minab, southern Iran. Bloomberg reported this month on an internal Pentagon review that has not been published.

The site was catalogued as an IRGC facility from outdated data. Satellite images from 2018 reportedly showed painted walls, a soccer pitch and playground markings. Staff reportedly expected Maven to flag stale records, which it was not built to do. Target preparation that took hours was reported to take minutes.

Senators Mark Warner, Jack Reed and Chris Coons wrote to Defense Secretary Pete Hegseth and DNI Jay Clayton. CNN's report, published 19 September, says it went out on the Friday, and a Warner release dates it 21 September. They asked for "immediate investigation by relevant Inspectors General" and for access to both cases.

The Pentagon's Minab review is unpublished. Its conclusions reach us through Bloomberg.

Allison Murphy, chief operating officer of Axios, spoke at ISOJ 2026, a journalism conference. The write-up of day two was published on 19 September. The line everyone quotes is the write-up's own sentence, since Murphy is not quoted directly: a news roundup section of a daily newsletter "that used to take reporters three hours can be completed in about 10 minutes with AI tools".

Read the scope. It is one section of one daily newsletter. Axios is also testing a "things to do" newsletter that the write-up calls "about 90% AI-generated". Nobody says what the 90 percent means. Is it text, links or selection?

Now the money. The Next Web reported on 17 August that OpenAI will fund the start of 13 new Axios Local newsletters over three years, covering staff and technology. Axios gives OpenAI training rights to its published content. The amount is not public.

I found no independent measurement of the time saved. I found no word on who checks the 10 minutes.

Zach Lloyd, chief executive of Warp, told Claire Vo on the How I AI podcast how Warp ships "2,000 PRs a month" with what he calls AI factories. That is the episode's title. Show pages give two numbers. It takes 35 minutes to go from a Slack message to a pull request with a QA video. It then takes 3.5 hours to reach the first human review.

Treat these as Warp's own figures. Warp sells the product, and the numbers come from a podcast interview. The two times probably measure different things, so the comparison is rhetorical. The pages do not define what counts as a PR.

A second voice in passing. Linear's engineering blog posted "CI bottleneck reworked" on 21 September, by Mufeez Amjad. It says: "Agents have made it exponentially faster to ship code, but validating those changes hasn't quite kept up at the same rate."

Linear's bottleneck is CI run time and cost, and Linear fixed it. That is a machine queue. Warp's claim is about a human queue. The two are not the same evidence.

arXiv describes itself as "a curated research-sharing platform open to anyone". Paul Ginsparg started it in 1991. Cornell took over its running in 2001. It is not peer reviewed: "Material is not peer-reviewed by arXiv." Submitting is free. In April it passed 3 million articles and reports more than 5 million active monthly users.

On 1 July it spun out of Cornell as arXiv, Inc., a nonprofit Delaware corporation with US tax exemption. The initial board has eight members. Cornell and the Simons Foundation are founding members for up to five years and named the first board. By my count, five of the eight come from Cornell or Simons.

On 23 September arXiv announced 17.2 million dollars in commitments over three to five years. The donors are Simons Foundation International, XTX Markets and the Siegel Family Endowment. The money goes to operations, to technology, "including work related to the management of AI-generated content", and to setting up the nonprofit.

Moderation is supported by volunteers. On 14 May the computer science section chair Thomas Dietterich presented a policy of a one-year ban for incontrovertible evidence of unreviewed AI text, according to The Next Web.

The gifts run three to five years from September 2026. They end somewhere between 2029 and 2031.

Thanks for reading this far. Mattias