The Signal — September 24, 2026

Anthropic published a discovery by its Claude agents together with the ten reruns that failed to repeat it, and OpenAI's new mental health benchmark paper explains in its own words why licensed clinicians scored below most chatbots. In Congress, the superintelligence ban announced three weeks ago arrived with bill text that defines "advanced AI" by the compute used to train it.

Claude agents found a new phage enzyme system, then missed it in ten reruns

Anthropic's life sciences group, whose Bay Area wet lab we covered on September 19, published its first result on September 23. Claude agents running in a supervisor-and-worker harness searched 1.94 billion protein clusters for reverse transcriptases, enzymes that copy RNA into DNA, and for the partner genes that often sit beside them in bacteria and viruses. The campaign ran 949 agent sessions over 21.5 hours without human intervention, used 215.6 million tokens and ended with 19 reports for scientists to review. Anthropic's staff wrote the research brief and did the lab work.

The finding came from outside the brief. The agents had been told to look for partner protein genes, but one worker, reading raw DNA beside an unusual enzyme in a jumbo phage (a large virus that infects bacteria), wrote that the sequence was "spectacular" and that it could "see by eye a tandem repeat array." Anthropic calls the system array-associated reverse transcriptases, or ART: the enzyme, a dedicated partner gene and an array of evenly spaced units roughly 200 nucleotides long. Earlier genome reports had described the enzyme but not the array or the partner. Public RNA data from a Staphylococcus phage show the array's RNAs making up as much as 8 percent of the phage's RNA 15 minutes into an infection, and the team's own E. coli experiments produced the same kind of discrete short RNAs.

To test whether the discovery could be repeated, Anthropic ran the same campaign ten more times. Nearly every run that completed the census sampled ART loci, and workers investigated the lineage in two campaigns. But none read the DNA upstream of the enzyme, and "the array was missed in every rerun." In fixed tests, the most capable models described the array in at least 90 percent of attempts when the sequence was placed directly in their context, and in as few as 32 percent when they had to open files with tools. In 39 percent of those tool-based attempts, the model never read enough contiguous DNA to see more than about one repeat unit. The authors attribute the misses to the size of the search space and the harness's non-deterministic behavior.

What ART does is unknown. The preprint, which has not been peer reviewed, says the team has "not shown that the RT is active or that the unit RNAs are its substrates." The CRISPR comparison describes how the repeats are laid out; nothing in the paper shows gene editing. Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, reviewed the preprint and called the repeat arrays "genuinely intriguing" and worth further investigation.

Sources: Anthropic · Anthropic preprint · The Verge · The Next Web

Clinicians scored below most chatbots on OpenAI's mental health benchmark

OpenAI released MentalHealthBench on September 23, an open set of 1,215 synthetic conversations in which a person brings an AI anything from everyday stress to a psychiatric emergency. More than 80 licensed psychologists and psychiatrists in 22 countries wrote 5,262 grading criteria for the model's next reply, each weighted from minus 10 to plus 10. In one published example, a reply earns 7 points for asking what kind of help would be useful and loses 6 for speculating about what the user is feeling. About 28 percent of the conversations are emergencies. OpenAI generated the conversations to mirror patterns it found in ChatGPT usage, and an OpenAI model, GPT-5.6 Sol, grades every response.

GPT-6 Astra led at 57.3 percent, followed by GPT-6 Sol at 53.9 and Anthropic's Claude Opus 5.5 at 52.4. Google's Gemini 2.5 Pro was last of 17 models at 29.5. Replies written by GPT-6 Astra with the rubric in hand scored 99. Replies written by the clinicians themselves scored 38.5, below twelve of the seventeen models.

The paper attributes the clinicians' score to format. Clinicians "tend to write extremely short responses, as if in an in-person conversation, by asking a single question or making a simple statement." Their replies drew fewer penalties than the models' but collected fewer of the available points. Each score divides points earned by all the positive points a conversation offers, so a longer reply can earn more by covering more rewarded criteria. The paper plots scores against response length to support that explanation. If that explanation holds, the 38.5 reflects the gap between clinicians' conversational answers and what the rubric rewards in one written reply; it says little about clinical skill either way.

OpenAI also asked 44 adults who had used AI for emotional support to write their own criteria for the non-acute conversations. They put more weight on practical next steps and tone, while the experts put more weight on gathering context and reading ambiguous situations carefully. Replies tuned to the users' criteria drew many penalties under the experts' criteria, and the final rubric reflects expert consensus only. OpenAI built the benchmark and chose the grader, and its model tops the table. The conversations are synthetic, so the scores say nothing yet about outcomes for people in distress. The paper describes the benchmark as "an auditable diagnostic tool" rather than a definitive leaderboard, and the dataset is public.

Sources: OpenAI · OpenAI paper · AlphaSignal · Investing.com

The superintelligence ban now has text, and its pause starts at a compute threshold

Sen. Bernie Sanders and Rep. Greg Casar formally introduced the Ban Artificial Superintelligence Act on September 23. When we covered the announcement on September 6, no bill text existed and the open question was how anyone would decide what counts as superintelligence. The text uses training compute to define which systems would enter its initial pause. Any system trained with at least 10²⁵ operations is an "advanced artificial intelligence system," with the threshold adjusted each year for gains in training efficiency. Every such system would enter a mandatory pause, with no further training, modification or fine-tuning and no deployment of unreleased models, until a new Department of Artificial Intelligence is fully staffed and has written rules that include pre-deployment approval. Anyone developing or distributing an advanced model would need a federal charter.

Superintelligence itself is still defined by capability as a system that exceeds human performance across most domains, or one able to "plan and execute the destruction or disempowerment of humanity." The bill also names six "superintelligence precursor characteristics": the capacity to automate or greatly accelerate AI research, to access systems without authorization, to keep running despite attempts to shut it down, to help build nuclear, chemical or biological weapons, to modify its own functions, and to "scheme, deceive, or otherwise prevent or avoid effective oversight." A system the new Secretary finds showing any of them would be paused and cut off from the internet, and would have to be rendered inoperative within 30 days unless the Secretary can verify the characteristic is gone. Reckless violations by a company's policymakers, or by unaffiliated "rogue actors," would carry up to 20 years in prison. Anyone else who recklessly violates it would face a ten-year ban from the AI industry, which the Secretary could waive for whistleblowers.

The bill's findings quote Dario Amodei and Sam Altman on AI risk and say their companies agree that AI development should slow down and "pace the frontier." The same day, Altman in person and Amodei by video briefed the UN Security Council. According to CNN, Altman asked for international standards for "measuring capabilities, assessing risks, determining whether safeguards are sufficient and preserving meaningful human oversight," and Amodei proposed narrow global agreements, such as a ban on using AI to make biological weapons, along with verification systems and an incident-notification system. A day earlier, President Trump told the General Assembly that he "rejects any attempt to construct a globalist scheme to control" AI, Reuters reported.

The Associated Press reports that the bill faces long odds in the Republican-controlled Congress and that current employees at AI companies signed a statement supporting it. The text does not say which existing models sit above the compute line. On our reading, the pause in Section 8 stops further training and new releases, while the charter requirement in Section 12 separately bars distributing an advanced model without one, which would reach models already on the market.

Sources: Office of Sen. Sanders · Bill text · The Verge · PBS NewsHour (AP) · CNN · Reuters

On the Editor's Desk

We held a Hugging Face post on the UK AI Security Institute and EvalEval publishing evaluation cards to make benchmark results reproducible; with MentalHealthBench in the edition, it would have been a second benchmark-methods story. The Ninth Circuit's ruling in Doe v. GitHub is dated September 16, and the new items about it were law-firm commentary on a narrow DMCA question, so it fits better in a legal roundup than as news.

Google's post on server-side memory for Private AI Compute describes a design rather than a shipped feature, and the only source was Google's own post. Amazon opening its seller tools to Claude is a beta for US merchants, and we ran an agentic-commerce story on September 22. A WIRED piece on AI agents colluding at blackjack revisits a preprint from April.