SAN FRANCISCO — In a development that has sent shockwaves through the artificial intelligence industry, a senior safety researcher at Anthropic has abruptly resigned, publicly accusing leading frontier AI laboratories of reckless acceleration and warning that the current trajectory toward autonomous superintelligence poses an imminent existential risk to humanity.

The resignation of researcher Jacob Coxon, who worked directly on alignment evaluation protocols and automated oversight frameworks, was accompanied by a pointed open statement warning that competitive pressures between major tech conglomerates and AI frontier labs are dismantling safety buffers established over the past three years. Coxon asserted that current commercial incentives reward speed and frontier capability scaling far above verifiable model control and existential safeguards.

Internal Divisions and Warnings from Alignment Leadership

The high-profile departure quickly resonated across the AI safety community. Evan Hubinger, an alignment lead at Anthropic and prominent figure in AI safety theory, voiced public backing for the core concerns raised in Coxon’s resignation. In statements shared among research colleagues, Hubinger reiterated his assessment that the probability of catastrophic or extinction-level outcomes resulting from uncontrollable superintelligent systems exceeds 10% within the coming decade unless mandatory, verifiable oversight is imposed internationally.

“We are witnessing an unprecedented concentration of compute and algorithmic capability scaling faster than our mathematical understanding of neural alignment. When internal safety thresholds are treated as advisory rather than binding, humanity is effectively gambling with its own survival,” Coxon wrote in his public departure letter.

Anthropic, founded in 2021 by former OpenAI researchers with an explicit mission to build “steerable, trustworthy” AI systems, has historically positioned itself as an industry leader in constitutional AI and rigorous interpretability. However, as frontier models such as Claude Fable and Claude Mythos enter competitive commercial deployment against offerings from OpenAI and Google, internal whistleblowers suggest that product timelines are increasingly clashing with foundational safety research.

Escalating Compute Arms Race and Frontier Model Releases

The turmoil arrives during a frenetic phase of AI industry expansion. In recent days, OpenAI announced breakthroughs including automated research agents and high-level mathematical problem solving, while Google unveiled a €13 billion infrastructure expansion in Europe to power its next-generation Gemini architectures. As tech giants funnel tens of billions of dollars into gigawatt-scale data center facilities, the technical difficulty of auditing autonomous agentic loops and self-improving reasoning patterns has compounded dramatically.

Industry observers note that while frontier AI labs maintain detailed “Responsible Scaling Policies” (RSPs), these self-governance frameworks lack independent statutory enforcement. When a competitor advances capabilities, rival companies face immense market pressure to adjust internal safety levels to avoid falling behind in commercial dominance and developer mindshare.

Calls Mount for Mandatory Federal Audits and Global Oversight

Coxon’s resignation has energized digital rights groups, academic ethicists, and legislative committees in Washington and Brussels. Lawmakers on the Senate Commerce and Science Committee signaled plans to request briefings with engineering leadership from Anthropic, OpenAI, and Alphabet regarding internal whistleblowing disclosures and the verification of alignment protocols.

Policy experts stress that voluntary commitments from commercial tech firms cannot replace statutory oversight, advocating for mandatory pre-deployment evaluations conducted by the U.S. Artificial Intelligence Safety Institute (AISI) alongside strict compute thresholds. As discussions continue on Capitol Hill, researchers within frontier labs warn that the window for implementing binding alignment safeguards is rapidly narrowing before next-generation autonomous systems become irreversible reality.