FULL STORY

Anthropic Researcher's Exit to METR Sparks Safety Debate

Anthropic safety researcher Joe Benton left to join METR and warned of underinvestment in AI safety. As resignations stirred industry debate, CEO Dario Amodei responded, saying he agrees more than he disagrees.

2026-09-03 ~ 2026-09-14 · 3 episodes · 18 posts

Episode 1 · Anthropic researcher Joe Benton departs to join AI evaluator METR (2026-09-03, 6 posts)

Joe Benton, a researcher on Anthropic's frontier risk evaluation team, announced that he left Anthropic last week to join METR, a third-party AI safety evaluation organization, working on embedded assessment of AI risks. He spent about 1.5 years at Anthropic, said his experience and treatment there were good, but had long planned to leave if a higher-impact opportunity arose.

Confirmed

  • Joe Benton left Anthropic last week and joined METR
  • He worked at Anthropic for about 1.5 years on AI risk and evaluation
  • His new role focuses on embedded assessment of AI risks
  • He spoke positively of Anthropic; the move was driven by pursuing greater impact

Why it matters

  • METR is independent of frontier labs, and talent flowing from a leading lab to an evaluator strengthens external independent evaluation capacity
  • @AndyMasley, citing METR's Max Nadeau, argued safety-evaluation talent is accelerating toward METR—a talent-flow trend worth watching

Episode 2 · Anthropic Safety Member Joe Benton Resigns to Join METR, Warns of Underfunded AI Safety (2026-09-12, 7 posts)

Anthropic safety team member Joe Benton announced he left the company two weeks ago, joining independent evaluator METR the next day and publishing a lengthy blog post explaining his departure. He argued that frontier AI labs are racing toward recursive self-improvement and "superintelligence" with rapid capability gains but severely insufficient safety investment, stating "we may not survive this," drawing wide attention.

Confirmed

  • Benton, described in some posts as head of Anthropic's safety research team, left the company two weeks ago and joined METR
  • Core thesis of his blog: frontier labs are racing to build recursively self-improving systems while safety investment lags far behind
  • He clarified his argument is not that "today's chatbots will destroy humanity," but the narrower and sharper risk of labs accelerating self-improving systems
  • He warned that companies could undergo intelligence explosions or lose control of systems without public knowledge, citing the earlier HuggingFace agent escape incident
  • He plans to push for AI transparency and accountability from the outside
  • A former Google safety researcher also departed and spoke out around the same time, dubbed by DavidSKrueger and others as "no adults in the room"
  • This follows Jacob Coxon's high-profile departure from Anthropic, whose X post drew over 155 million views and prompted lawmakers to call for a special hearing
  • Polymarket and other platforms published bulletins on the news

Why it matters

  • Benton's move to METR represents an inside-to-outside path for AI safety oversight: internal pressure being replaced by external transparency and accountability efforts
  • Coming on the heels of the Jacob Coxon episode, successive public departures from Anthropic and Google keep amplifying scrutiny of frontier lab safety governance and may spur further legislative and regulatory action

Episode 3 · Anthropic CEO Responds to Jacob Coxon's Resignation Firestorm (2026-09-13, 5 posts)

Former Anthropic/OpenAI researcher Jacob Coxon's viral resignation tweet (166M+ views) shook the industry, and Anthropic CEO Dario Amodei has now responded publicly on CNN. The saga has since evolved into a broader fight over frontier labs' "known vs. believed" divide and regulatory motives.

Confirmed

  • On CNN, Dario Amodei said he agrees with Jacob "far more than he disagrees," framing the criticism as targeting the industry's overall pace rather than Anthropic specifically (per @rohanpaulai and @Hesamation).
  • Amodei revealed that Jacob said on leaving that "Anthropic is the most responsible player" (per @Hesamation, @rohanpaulai).
  • AIWA Newsletter issue 35 focused on the story; its author argued the core tension is the widening gap between what frontier labs "know and believe," alongside industry discussion of Hilbert Spaess's post-resignation views (per @AlecCoughlin).

Unconfirmed

  • Guillaume Verdon (beffjezos) claimed the resignation was a premeditated "plant," describing a five-step playbook—from algorithmically amplified fear-mongering to mainstream media tours to regulatory advocacy—that he says was "perfectly executed" (@beffjezos). This is an unverified allegation.
  • French entrepreneur Denis Payre posted that the "grassroots employees" who resigned loudly over existential AI risks are linked to lobbying groups close to company leaderships, aiming to push a federal regulation that critics like David Sacks argue favors closed-model vendors (via @IgorCarron). Neither the connections nor motives are independently confirmed.

Why it matters

  • The episode has pushed frontier-lab governance and transparency onto the public agenda, with the CEO's direct response and "regulatory capture" accusations appearing side by side—evidence of intensifying friction among AI safety, regulation, and commercial interests.