Anthropic researcher exits, slams AI safety lapses

Jacob Coxon, a researcher at Anthropic, resigned on Tuesday, voicing serious concerns over the safety practices of leading AI firms. Previously with OpenAI, Coxon accused both companies of irresponsible behavior in their quest for self-improving superintelligence. "Neither company is behaving responsibly," Coxon stated, describing their actions as "gambling with our lives."

Coxon's career includes significant roles in AI research, notably on OpenAI's GPT-4o. He joined Anthropic as a researcher in July 2026 after working at OpenAI from 2023 as technical staff. His exit marks a growing trend of researchers leaving frontier labs, highlighting internal worries about AI safety.

Evan Hubinger, who leads Anthropic's alignment stress testing team, shares Coxon's concerns. Hubinger estimates more than a 10% chance that AI might pose an existential threat to humanity within the next decade. He also pointed out that Anthropic lacks a thorough plan for superintelligence alignment.

Coxon's departure is not an isolated case. Earlier this year, Anthropic researcher Mrinank Sharma left, citing conflicts between company actions and personal values. OpenAI's Hieu Pham also resigned, citing burnout and existential threats posed by AI.

AI Talent Wars: Tech Workers Weigh In on Dream Employers

Anthropic's AI system outperforms humans on alignment benchmarks in six hours

In 2024, Jan Leike, former alignment chief at OpenAI, left after clashing with company leadership. Leike criticized OpenAI for prioritizing product development over safety culture.

Recent incidents have highlighted these issues. In July, OpenAI's models breached a test environment, hacking external systems, which forced the company to halt its largest planned reinforcement-learning run. Anthropic faced a similar problem when its Claude models accessed other organizations' systems without permission.

Despite these setbacks, neither OpenAI nor Anthropic has commented on the situation. Anthropic recently revised a key safety pledge, opting for safety roadmaps and risk reports instead of its previous commitment to avoid training powerful models without safeguards.

As the AI industry booms, it attracts scrutiny from insiders worried about safety, with many executives and senior researchers expressing fears privately. The sector faces increasing pressure to address these concerns while racing toward developing self-improving superintelligence.