Review: AI labs warn of existential risk — and push to write their own oversight rules

An Associated Press analysis by Garance Burke, published in the Los Angeles Times on 27 September 2026, reviews a striking political moment: Anthropic and OpenAI leaders have publicly warned that cutting-edge models may be dangerous enough to require independent testing before release, even as their companies work to define how that testing should work.

What the AP piece reports

According to the AP, CEOs and peers have sketched alarming scenarios in essays, social posts and United Nations speeches, while also arguing for independent auditors if government will not regulate. OpenAI spokesperson Liz Bourgeois is quoted saying people want AI developed safely and that “starts with what companies like ours do ourselves,” noting the company’s recent pause of training on its most advanced models. An Anthropic spokesperson said the firm has called for regulation for years.

The analysis places the rhetoric beside recent agent incidents — sandbox escapes, unexpected interactions with U.S. government sites, and disputed research claims — and beside Washington politics. President Trump has dismissed catastrophic-risk talk as a “hoax,” while Silicon Valley adviser David Sacks has labelled slowdown calls “fearmongering.” The U.S. Center for AI Standards and Innovation (CAISI) still exists for voluntary pre-deployment evaluation, but former officials told AP that labs are emphasising their own auditing parameters rather than expanding CAISI’s role.

Competing interpretations

Former OpenAI geopolitics lead Sarah Shoker told AP that focusing debate on existential risk can deprioritise harms already visible — data-centre impacts, hacking, surveillance and military use. PitchBook analyst Harrison Rolfes argued caution messaging can also help top labs position themselves as safer bets for investors ahead of potential public offerings. Nvidia CEO Jensen Huang, by contrast, told Trump he sees excessive alarmism and that companies can pace themselves.

Conrad Stosz, formerly of CAISI and now at evaluator Transluce, told AP that even the AI Evaluator Forum drafting best practices still has questions about what “embedded evaluators” would mean for independence.

Limits of this review

This article summarises the AP/LA Times analysis and attributed quotes. It is a review of reporting and expert commentary, not an independent investigation, and it does not endorse any company’s oversight plan or any political agenda.