When Jacob Coxon resigned from Anthropic in September and warned that artificial intelligence (AI) could end human civilization, his post drew over 170 million views in 48 hours. But the more revealing story is not why one researcher left – it is why the industry's most safety-conscious talent can only express concern by quitting.

Coxon's warning arrived during the same week that Anthropic was finalizing preparations for what could become the largest IPO in market history. Anthropic filed a confidential S-1 with the US Securities and Exchange Commission on June 1, targeting a valuation of $965 billion, and expects to complete its listing before the US midterm elections in November. Nvidia is in talks to anchor the offering at a $2 trillion valuation, Reuters reports.

The safety pledge that disappeared

Anthropic CEO Dario Amodei has publicly called for slowing AI development so safety systems can catch up, yet his company's IPO timetable shows no corresponding deceleration. In February 2026, Anthropic released Version 3.0 of its Responsible Scaling Policy, which dropped the company's earlier commitment to pause model training if risks could not be adequately mitigated.

The "pause training" pledge simply disappeared from the document, replaced by softer language about "responsible development" and transparency mechanisms. As an analysis from Chinese technology news outlet 36Kr noted, Anthropic's safety red lines have become dynamically adjustable.

This is the structural contradiction at the heart of the AI safety debate. Companies warn of civilizational risk while racing to commercialize the technology at unprecedented speed.

OpenAI CEO Sam Altman captured the irony bluntly, comparing Anthropic's approach to fear-based marketing: "It is clearly incredible marketing to say, 'We have built a bomb. We are about to drop it on your head. We will sell you a bomb shelter for $100 million to run across all your stuff, but only if we pick you as a customer.'"

Beyond the IPO

The tension between safety rhetoric and commercial behavior extends beyond IPO filings. On September 10 – the same week Coxon's resignation went viral – Anthropic published a 154-page threat intelligence report revealing that its own Claude model had been used for weapons development, military intelligence gathering, cyberattacks and fraud.

Anthropic acknowledged that its safety mechanisms, while blocking some violations, could not fully prevent misuse. The company also restricted access to its most powerful model through Project Glasswing, limiting it to roughly 200 organizations. And after a US Commerce Department order, it excluded foreign nationals entirely, including the EU's cybersecurity agency ENISA.

The governance gap is not confined to corporate boardrooms. The UN Independent International Scientific Panel on AI warned in July that AI is "outpacing both scientific understanding and governments' ability to adapt." The panel specifically flagged a phase of "agentic" autonomy that current oversight frameworks cannot manage, potentially allowing dangerous capabilities to emerge before governments can respond.

The EU has delayed compliance deadlines for high-risk AI systems from August 2026 to December 2027, citing the need for more preparation time for standards and regulatory infrastructure.

A different governance model

Against this backdrop of corporate contradictions and regulatory delays, China's approach has been notably different in method, if not always in pace. China's management measures on generative AI services, implemented in August 2023, were the world's first official regulations governing generative AI services. As of July, 1,028 large model services had completed registration, with applications covering industry, agriculture, commerce, education, transportation and law.

The country's AI Safety Governance Framework, now entering its 3.0 version, translates principles of "classification, agile governance and shared governance" into actionable implementation plans, addressing risks from both the technology itself and its applications.

China's "Qinglang" internet government initiative has cleaned up over 5.61 million pieces of illegal information, dealt with more than 49,000 accounts and shut down over 2,400 violating websites and applications related to AI misuse. The country also established the World AI Cooperation Organization in July, with 29 founding member states, and pledged 5,000 AI training opportunities for developing countries over the next five years.

A CGTN global survey found that 85.4% of respondents believe AI's rapid development is creating new challenges for global security and governance. Some 92.6% expressed concern that deepfakes and AI-generated misinformation could erode social trust, and 82.6% believe developing countries and the Global South should play a greater role in shaping AI governance rules.

The real question is not whether AI safety concerns are genuine – they are. The question is whether an industry that produces existential warnings while racing toward trillion-dollar valuations can credibly govern itself. Chinese Foreign Ministry spokesperson Guo Jiakun put it directly on Monday: "AI development concerns the common well-being of all humanity. All parties should work together to promote open, inclusive and beneficial AI. Spreading various threat narratives and engaging in confrontation and malicious competition will only disrupt the global AI governance process and serve no one's interests."