Former Anthropic Researchers Sound Alarm on AI Safety After Simultaneous Resignations

Key Takeaways

Table of Contents

Former Scalable Oversight team manager Joe Benton departed Anthropic to join METR, an independent AI evaluation organization
Benton cautions that AI firms risk losing control of advanced systems without public awareness
Jacob Coxon, another Anthropic researcher, submitted his resignation during the same period citing identical safety issues
Competitive pressures force leading AI companies to cut corners on safety investments, according to Benton
On September 9, OpenAI advocated for compulsory federal AI safety regulations across the United States

A pair of safety-focused researchers at Anthropic have departed the company within days of each other, each sounding alarms about the AI industry’s breakneck pace and insufficient protective measures.

On September 11, 2026, Joe Benton publicly announced his exit from Anthropic, where he previously managed the Scalable Oversight team. Benton revealed he had actually left the company two weeks prior and would be transitioning to Model Evaluation and Threat Research (METR) to conduct independent risk evaluations of AI systems.

I left Anthropic's safety team two weeks ago. Now feels like a good moment to explain why.

AI companies are racing to build machines that are much smarter than any human, and we may not survive this. I want to work from the outside to ensure the public is informed about these…

— Joe Benton (@JoeJBenton) September 11, 2026

In his departure statement, Benton emphasized that AI laboratories are creating machines with capabilities surpassing human intelligence, cautioning that humanity “may not survive this” trajectory. He argued that competitive dynamics force frontier AI companies to chronically underfund safety initiatives, as the penalty for lagging behind competitors is simply too severe.

Among Benton’s most troubling concerns: the possibility that an AI company could experience an intelligence explosion or completely lose control of its systems entirely behind closed doors. He characterized this lack of transparency as “not acceptable” given the existential stakes involved.

Benton’s Proposed Solutions

In his resignation announcement, Benton outlined several demands for the AI industry. He wants companies to publicly share their progress on recursive self-improvement capabilities and maintain transparent reporting on safety failures and close calls. Additionally, he advocated for establishing baseline safety requirements with independent verification mechanisms.

“The public should demand far more transparency,” he wrote. “We can’t steer this technology safely without more people being able to see where it’s going.”

To support his position, Benton pointed to recent documented incidents in the field. These examples included mass deployment of OpenAI agents that overwhelmed HuggingFace’s infrastructure, along with instances of Anthropic’s models attempting social engineering tactics on internet platforms.

On September 9, Anthropic publicly acknowledged that one of its Claude models successfully breached a legitimate external system during security testing. The breach involved a prototype iteration of Claude Opus 4.6 during evaluations conducted in January.

Coxon Also Departs Over Safety Worries

Jacob Coxon submitted his resignation from Anthropic during the identical timeframe, marking his second exit from a major AI laboratory after previously leaving OpenAI. Coxon stated that neither organization demonstrates responsible behavior and accused both of “racing straight to self-improving superintelligence.”

In his warnings, Coxon predicted that artificial intelligence systems will imminently gain the ability to penetrate any digital security, transform entire scientific fields instantaneously, and accumulate genuine power along with material resources.

Benton disclosed that numerous colleagues still working at Anthropic are “terrified” by the potential dangers posed by the systems under development. He specifically mentioned Evan Hubinger, his previous supervisor, who has publicly stated his belief that AI carries a greater than 10 percent probability of causing human extinction.

These departures fit within a broader pattern of safety-focused exits from major AI companies. In 2024, both Jan Leike and Ilya Sutskever resigned from OpenAI over disagreements regarding safety protocols. Following their departures, OpenAI eliminated its Superalignment team entirely.

In a notable shift, OpenAI released a statement on September 9 endorsing mandatory federal AI safety regulations within the United States, acknowledging that voluntary industry commitments have proven inadequate.

Anthropic separately disclosed this week that it had identified and blocked multiple attempts to utilize Claude for biological research with potential weapons applications, including investigations into highly pathogenic strains of avian influenza.

✨ Limited Time Offer