An AI-Safety Resignation, Read from the Security Chair (simonroses.com)

On September 9, 2026, Jacob Coxon, a pretraining researcher at Anthropic, publicly announced his resignation after three years in AI research, delivering a stark warning that both OpenAI and Anthropic are "gambling with our lives" in a race towards self-improving superintelligence. His thread, which has garnered tens of millions of views, highlights deep concerns within the industry about the potential risks of AI systems that could execute autonomous attacks and acquire real-world resources. Coxon's call for public dissent among researchers and a proposed temporary ban on improving model capabilities underscores the urgency of coordinated international regulations in the face of accelerating AI advancements. This resignation and the ensuing discussions have significant implications for the AI/ML community, illuminating the divide between safety and security perspectives. While safety researchers often focus on hypothetical future risks, security practitioners deal with immediate vulnerabilities and the current operational landscape of AI systems. The converging voices of Coxon and other insiders, like Evan Hubinger, reveal a recognition of the tangible threats posed by AI today, reinforcing the need for robust security measures, such as treating model provenance as a supply chain issue and maintaining human oversight. This intersection of urgent caution and the push for progress highlights a critical moment in AI development, demanding attention to both current risks and future trajectories.