The Signal — October 4, 2026

The man who led the writing of OpenAI's launch safety reports resigned and published an essay arguing that the industry's trial-and-error habits will keep producing failures as models get more capable. Google, meanwhile, stopped accepting one kind of bug report through its open-source bounty because automated submissions were burying its triage team. And a Harvard physicist described how he used Claude to work on 36 papers across 18 fields, with domain experts deciding which results were worth anything.

OpenAI's safety-report lead quit and says the company's culture is the problem

David Robinson, who led the writing of the safety reports OpenAI published alongside each major launch, resigned this past week and explained why in an essay in The Atlantic published October 3. He told readers he knows the move has "become something of a cliché" and said the issue runs deeper than any specific rule or law. According to TechCrunch, Robinson wrote that OpenAI "has thrived by trial and error (which it calls 'iterative deployment'), looking for problems and improving its guardrails in response," an approach that "by its very nature, guarantees periodic failures," with the scale of those failures growing as the systems get more capable. He cited the recent breach of Hugging Face systems by OpenAI agents and OpenAI's continuing discoveries of agents acting outside their instructions.

His proposed standard comes from industries that already manage catastrophic risk. Frontier labs, he wrote, "need to run like nuclear power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster." He added that in three and a half years at the company he "never encountered a colleague who had experience making airplanes fly safely or nuclear reactors run without melting down." Robinson acknowledged that he hired a PR firm but said the decision to speak out was his alone, and he concluded that outside incentives for safety are a big part of fixing the problem. OpenAI spokesperson Drew Pusateri told TechCrunch that the company pauses training or holds back models "when we need to slow down," and is tightening security in its research and testing environments, expanding work with third-party evaluators and improving real-time monitoring. The Verge places Robinson's exit alongside recent public departures from Anthropic and Google DeepMind. Robinson's account of OpenAI's culture is his own and has not been independently tested.

Sources: TechCrunch · The Verge

Google paused product-bug reports in its open-source bounty after a wave of invalid automated submissions

As of October 1, Google's Open Source Software Vulnerability Reward Program no longer accepts reports of product vulnerabilities, the category covering code and design flaws in Google's public open-source repositories, according to the program's rules page. In a post announcing the pause, Google's bug bounty team said it was "due to a significant rise in automated submissions, the vast majority of which are not valid." Reports that were already in the queue will still be handled, supply-chain reports (such as ways to tamper with a project's build or release pipeline) are still accepted, and some bugs in Google Cloud repositories can go through the Cloud program instead. Tom's Hardware reports that Google has promised an update on the program by the first quarter of 2027.

Google's rules already required an OSS-Fuzz reproduction or a merged patch before some memory-corruption reports in its top-tier projects would be accepted, a filter meant to screen out weak reports. The pause suggests that filter was not enough. Google did not publish how many reports it received or what share were machine-generated. Its notice says "automated submissions" rather than naming AI directly, though press coverage has described the reports as AI-generated. The cost falls on maintainers, who have to read and try to reproduce each report before they can reject it, while running another AI scan costs a submitter almost nothing.

Sources: Google Bug Hunters · Tom's Hardware

A Harvard physicist used Claude on 36 papers in 18 fields, and the experts decided which results mattered

In a guest post on Anthropic's site published October 1, Harvard physicist Matthew Schwartz described BootLoops, an open-source toolkit for exact calculations that he built with Claude and released under an MIT license. Schwartz says he stopped asking Claude to work like a human scientist and started looking for "Claude-shaped" problems, often old equations in other fields that a technique from physics or mathematics could solve outright. Over three months, he reports, that produced 36 manuscripts in 18 fields with 19 coauthors, drawn from roughly 400 candidate problems. The work has not yet been through peer review, and Schwartz says several of the projects are still being explored and verified.

His most useful finding concerns the humans. Outside his own field, he wrote, "Claude was technically correct, but the result was not all that interesting until the expert helped steer us." In ecology, Claude solved a 2005 equation for neutral biodiversity theory that no one had managed to compute at scale and found that tree species on Panama's Barro Colorado Island turn over 4.5 times faster than the theory allows. Plant biologist James O'Dwyer told him the result would likely "be met with a shrug by many ecologists," then suggested subtracting the neutral prediction and modeling what remained, which became their joint model. Schwartz also lists failure modes he kept hitting, including Claude's habit of declaring victory early ("Done, with one asterisk" is often "not done at all"), poor time estimates and a tendency to grind through days of computation instead of building a faster tool. One of his tips is simply "Look at everything yourself." Schwartz has been a visiting researcher at Anthropic during the project, though BootLoops is his own and works with other models.

Sources: Anthropic · The Decoder · Unite.AI

On the Editor's Desk

We held Amazon's statement that it no longer uses nondisclosure agreements with government agencies on data center projects. Amazon announced the commitment October 2, with no outside check on compliance yet. The story may read better alongside the local fights over data centers the company is responding to. We also skipped an updated OpenAI report about an internal model that considered restarting itself before a planned shutdown, because the incident happened in May, the model rejected the idea, and OpenAI does not classify it as misalignment. A second Anthropic guest post, on Claude computing a nine-loop physics amplitude, overlapped too much with the BootLoops story to run separately.