Rahul Somvanshi
Jacob Coxon-openai-ai-threat-jacob-coxon-b3047033.html), a researcher at Anthropic, announced his resignation from the company, stating that Anthropic and OpenAI are engaged in a race toward self-improving superintelligence while gambling with human survival. Coxon previously worked at OpenAI developing AI models and has expertise in model training. He contended that the technology could kill all humans by the end of the decade.
Evan Hubinger, another Anthropic researcher, responded publicly on X (formerly Twitter) that he personally calculates greater than a 10 percent probability that artificial intelligence could kill all humans within the next decade. He acknowledged that Anthropic is attempting to address alignment—the problem of ensuring powerful AI systems behave as intended—but stated the company has no plan to solve alignment at superintelligence scale.
The Financial Times reported that Anthropic had withheld its latest model from the UK AI Safety Institute, raising questions about industry cooperation with government safety oversight.
These statements emerged amid broader government engagement with AI safety issues. Prime Minister Andy Burnham acknowledged AI national security risks in Parliament and noted that AI systems can also provide solutions to safety challenges. The government appointed an AI safety minister, Kanishka Narayan, and granted this position a Cabinet table seat—elevating AI governance to high governmental priority.
Former minister Darren Jones called for international treaty frameworks to regulate development of superintelligent systems safely, implicitly acknowledging that national regulation alone cannot manage technologies with potentially global effects.
The UK government’s AI Safety Institute (AISI) tested OpenAI’s GPT-6 Astra model before its public release, according to the Cabinet Office. This represents one mechanism through which government attempts to exercise oversight of capabilities before models reach the broader public.
What these developments collectively illustrate is genuine tension within the AI research community about whether current development practices are sustainable and safe.
On one side, researchers like Coxon and Hubinger raise concerns about existential risk. They argue that creating artificial superintelligence without solved alignment problems is reckless. They point to probability estimates finding significant risk of catastrophic outcomes.
On the other side, AI developers argue that they are implementing safety measures and that cooperation with government bodies like the AISI represents responsible governance. They contend that advancement cannot be halted, only managed.
The core disagreement is not whether superintelligent AI poses risks—participants on both sides accept this premise. The disagreement is about whether current risk mitigation is adequate and whether capability development should proceed at current pace.
Coxon’s resignation and Hubinger’s public statements indicate that internal Anthropic debate has surfaced to public view. Researchers willing to leave positions and speak publicly about risks represent a significant signal that concerns are substantial enough to overcome career considerations.
The government’s appointment of Narayan and AISI testing of models before release find Westminster is taking AI safety seriously. Whether these institutions can impose constraints on private companies pursuing commercial AI development remains uncertain. The alignment problem—creating superintelligent systems that do what humans intend—remains unsolved across the industry.
**Word count: 416**
