The Signal — September 10, 2026

A researcher at Anthropic quit on September 8 and said in public what people inside these labs usually say only in private, and within about thirty hours two institutions had answered him: OpenAI put an alignment researcher on the board committee that oversees model safety, and California's governor signed two bills whose whole purpose is to let somebody from outside check the work.

An Anthropic researcher resigned in public, his colleagues agreed with him, and the next evening OpenAI seated Paul Christiano on its safety committee

Jacob Coxon spent about three years doing pretraining research, first at OpenAI and then at Anthropic, and four months into the Anthropic job he resigned and posted why. "Neither company is acting responsibly," he wrote. "They are racing straight to self-improving superintelligence and gambling with our lives." He is 27. Axios reported that he left two months before his Anthropic equity would have vested and that he still holds equity in OpenAI, which is worth stating up front rather than leaving for someone else to find.

What made the post travel was not the resignation. It was the reply. Evan Hubinger, who leads alignment science at Anthropic and still works there, wrote: "Jacob is correct here, we really do earnestly believe AI could kill all humans! I personally think it is greater than 10 percent within the next decade." He added that Anthropic does not yet have a plan to solve alignment for superintelligence and is not clearly on track to get one. That is a current employee, under his own name, confirming a departing colleague's account of what the building believes.

The following evening OpenAI announced that Paul Christiano is joining the OpenAI Foundation board and its Safety and Security Committee, chaired by Zico Kolter. The Foundation describes the committee as providing governance over safety and security practices across all of OpenAI; TechCrunch puts it more sharply, reporting that the committee has the final say on whether new models are released. Christiano founded the Alignment Research Center, led alignment research at OpenAI from 2017 to 2021, did foundational work on reinforcement learning from human feedback, and is currently a Senior Tech Advisor at NIST's Center for AI Standards and Innovation. He will be a non-voting observer on the OpenAI Group PBC board. In his own post he wrote that he now believes there is "a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term," and that the industry, OpenAI included, is not on track to reduce that risk to an acceptable level. He is joining anyway, on the theory that if OpenAI rises to the occasion the risk drops a great deal.

Two things are worth keeping straight. These are personal judgments from three people, not measured probabilities and not company positions; Hubinger's ten percent is a number he arrived at himself, and Christiano's read on the industry's trajectory is offered as he takes a governance seat inside one of the companies he is describing. And there is a recusal attached. The Foundation's announcement carries a footnote saying that in his NIST role Christiano will step back from all OpenAI-related matters and all model evaluations, so he is trading the government's evaluation chair for the Foundation's oversight chair rather than occupying both. Which of those two seats does more good is a genuinely open question, and he has now answered it for himself.

The uncomfortable version of this week is that the people with the most detailed picture of these systems keep describing a risk they cannot bound, and then keep going. Coxon's answer was to leave. Christiano's was to take a seat inside. Hubinger's, so far, is to stay and say so out loud. None of those is obviously wrong, which is most of the problem.

Sources: TechCrunch · OpenAI Foundation · NPR · Mashable

Newsom signed two bills that build a profession of outside AI auditors

On September 9 Governor Gavin Newsom signed SB 813 and AB 1405, both of which show as chaptered in the legislature's records the same day. Neither touches what a model is allowed to do. SB 813, from Senator Jerry McNerney, creates a category called an independent verification organization and tasks the state with defining the criteria for one and checking whether a given outfit actually meets them. AB 1405, from Assemblymember Rebecca Bauer-Kahan, builds a state registry of AI auditors with standards for their independence and transparency. California already has laws telling frontier developers what to disclose. These two are about who gets to say whether the disclosure is true.

The politics around the signing were unusually visible. Anthropic backed the package in August. OpenAI came out for it on the ninth, hours before Newsom signed, with chief global affairs officer Chris Lehane saying the company intends to keep working state legislatures until Congress passes something national. Newsom's statement cited "the concerns raised in recent incidents," which is a governor's way of pointing at Coxon's resignation without naming it, and then called on the federal government to write real rules. McNerney was blunter, saying Washington is "unable or unwilling" to assess AI safety risks. Bauer-Kahan put the case for her bill in one line: we cannot expect industry to grade its own homework.

We wrote on September 5 about a different California bill, SB 947, the No Robo Bosses Act, which was then sitting on Newsom's desk and is still there, enrolled since September 4. That one would constrain what employers may do with automated decision systems. These two do something structurally different and, over a longer horizon, possibly more consequential. A disclosure regime with no independent auditor is a regime where the only people equipped to check the filings are the people who wrote them, and building the auditing profession is the unglamorous prerequisite for every rule that comes after it.

It is also almost entirely unbuilt. The bills create a registry and a credentialing process; they do not create people qualified to sit in that registry, or working methods for evaluating a frontier model from outside, or any obligation on a developer to hire one. "First-in-the-nation" is the state's own phrase, and on the narrow question of an auditor registry it appears to be accurate. Whether it amounts to oversight depends on who signs up, what they are actually permitted to inspect, and whether anyone has to listen to them.

Sources: Office of the Governor · POLITICO · California Legislative Information (SB 813) · California Legislative Information (AB 1405)

DeepMind precomputed a prediction for every single-letter change the human genome can make

We held this one yesterday because it landed next to a story about machines doing mathematics and deserved better than to be the third item. Google DeepMind released AlphaGenome Atlas on September 8: predicted molecular effects for roughly nine billion single-nucleotide variants, which is every possible one-letter substitution in the human genome, precomputed and published free for academic research. It runs to about a petabyte, more than thirty times the AlphaFold Database, and it is reachable through a web portal and the AlphaGenome API.

The piece that changes how the thing gets used is the AlphaGenome Variant Impact score. AVI folds AlphaGenome's regulatory predictions together with AlphaMissense's protein-impact predictions into one number per variant, so a researcher can rank nine billion possibilities and then read off which molecular process is predicted to break. That matters most in the 98 percent of the genome that does not code for protein, where most trait-associated variants live and where ranking has always been hardest.

DeepMind published two collaborator results alongside the release. Laura Covill and Anne O'Donnell-Luria at the Broad Institute used AVI ranking to surface a variant in DNM1, a gene linked to epileptic encephalopathy, that earlier analysis had passed over; the prediction said it created an incorrect splice site, and experimental screens confirmed it. Gareth Hawkes at Exeter ran Atlas against whole-genome data from more than 54,000 UK Biobank participants, grouped rare variants by predicted molecular effect, and recovered 22 percent more non-coding associations than were visible in the statistical noise otherwise.

Every number in the Atlas is a model output. DeepMind says so directly, and its own disclaimer states that AlphaGenome has not been validated or approved for any clinical use. The Broad result was checked at the bench; the Exeter number is a statistical finding nobody outside the collaboration has replicated yet. The honest description of what shipped is a very large, very fast hypothesis generator, free to academics, with an explicit warning not to mistake it for an answer. That is a useful thing to be, and the field has spent years wishing for exactly this while being unable to run the experiments.

Sources: Google DeepMind · MarkTechPost

On the Editor's Desk

Anthropic published an economic scenario explorer this week, with a technical report by Korinek and colleagues, laying out three conditional futures for the US economy through 2030. In the extreme one, output doubles every four and a half years and knowledge-worker unemployment reaches 17.9 percent, roughly where Dario Amodei's own warnings from last May land. The paper assigns no probabilities to any of the three. It is worth its own space, and running it today would have made two thirds of this edition about Anthropic arguing with itself, so it is queued rather than dropped.

We passed on the Justice Department's inquiry into how Nvidia structured its 2025 Groq licensing deal. The New York Times has it and Reuters and Bloomberg followed, but it is all sourced to people who would not be named, with no filing or docket to point a reader at. We also skipped Google's Mantis, an open-source harness that lets coding agents find and patch vulnerabilities, because Google documents it as a demonstration rather than a supported tool, and the interesting part will be what it finds once somebody outside Google runs it on real code. The FDA's generative-AI device docket came back around in aggregator coverage with its October 19 comment deadline attached; we ran that story when the docket opened in August and the deadline is not new.