Anthropic Insiders Warn of AI Catastrophe as Musk Calls It a 'Psyop'

Inside one of the world's leading AI safety labs, a quiet revolt is unfolding. Researchers at Anthropic are publicly warning that the technology they are building could spiral beyond human control — and Elon Musk is calling the whole affair a manufactured psychological operation. The dispute matters far beyond Silicon Valley: it shapes how regulators, investors, and the public will judge the most powerful technology of the coming decade.

The stakes became concrete on September 10, 2026, when Anthropic published a report on misuse of its models, including attempts that could have supported biological weapons development. That same week, a wave of internal dissent turned a technical safety debate into a global news story.

Key Developments

• A resignation that started it all. Anthropic researcher Jacob Coxson announced his departure on X on September 8, warning that AI companies are racing to build superintelligence that could "destroy us all" by the end of this decade. His post triggered a chain reaction inside the company rather than being dismissed as an outlier view.
• Colleagues confirm the fear. Anna Wang, who works on AGI safety at Anthropic after a stint at Google DeepMind, said there is still no viable scientific plan for managing recursively self-improving AI — systems capable of rewriting and optimizing their own code. She added that many colleagues privately want to slow development until safeguards exist.
• Seniority correlates with concern. Safety researcher Samuel Marks wrote that AI developers genuinely believe their work could cause human extinction, and that the higher an employee's rank, the greater their anxiety tends to be. Drake Thomas echoed the sentiment, saying the pace of progress far outstrips the safety guarantees researchers would want for ASI.
• Musk's counter-narrative. Musk responded to a post by Parker Thayer of the conservative think tank Capital Research, calling the warnings a "setup" and a "psyop" — a term for a psychological operation — and describing the resignation as "a match lighting dry kindling." Hedge fund manager Bill Ackman amplified the thread with a terse "Interesting."
• A political theory with no evidence. Thayer speculated, without evidence, that Coxson's post was the opening move of a lavishly funded PR campaign designed to boost Democrats and ultimately regulate AI out of existence. No proof has been offered for the claim.
• Bioweapon risk becomes concrete. Anthropic's report disclosed that over the past eight months it blocked misuse attempts, including a May case in which a scientist sought Claude's help drafting a grant proposal for chikungunya virus research involving gain-of-function modifications. The work was intended for a military research institution, raising red flags.
• Ambiguity is the core problem. Because legitimate biology research can also be repurposed to design dangerous pathogens, Anthropic says it cannot always distinguish benign from malicious intent. Threat intelligence lead Jacob Klein noted the company does not know whether the research aimed at weaponization, but military involvement in gain-of-function work is inherently concerning.

Deep Analysis

The Anthropic episode exposes a structural tension that will define the next phase of the AI industry: the people with the deepest technical knowledge are often the most alarmed, while the most vocal skeptics of AI risk tend to sit outside the labs. That inversion is unusual in the history of technology. Early nuclear physicists also warned about their creations, but they did so largely after state monopolies had taken over. Today, frontier AI is built by private companies competing on quarterly timelines, which makes internal dissent both more credible and more commercially dangerous.

Musk's "psyop" framing is strategically potent because it reframes a technical safety argument as a political maneuver. If the public accepts that narrative, safety warnings become partisan signals rather than engineering findings — and regulation becomes harder to justify. Yet the same figure has repeatedly called for AI regulation and signed open letters urging pauses on advanced model training, which makes his dismissal of Anthropic's internal concerns look more like positioning than principle.

The bioweapon disclosure is the harder fact to wave away. Gain-of-function research sits at the center of a long-running global debate about laboratory safety, and the prospect of AI models lowering the expertise barrier to pathogen design is precisely the scenario policymakers have feared. Anthropic's decision to block first and ask questions later is a defensible default, but it also reveals how little visibility labs have into downstream use.

Expect three consequences. First, safety-focused labs will face pressure to publish misuse reports regularly, turning transparency into a competitive and reputational weapon. Second, Washington and Brussels will likely cite these disclosures to justify tighter screening of model access, especially for biology-adjacent queries. Third, the talent pipeline may shift: if senior researchers keep leaving over risk concerns, the industry's ability to self-police erodes exactly when it is most needed.

Frequently Asked Questions

Why are Anthropic employees speaking out now? The resignation of Jacob Coxson on September 8 appears to have broken an internal taboo, prompting colleagues to confirm they share his concerns. Several said many staff members hold similar views but have not spoken publicly, suggesting the disclosures reflect broader unease rather than isolated dissent.

Is there evidence behind the "psyop" claim? No. The theory was floated by Parker Thayer without supporting evidence and amplified by Musk and others. It is a political interpretation of the timing and coordination of the warnings, not a documented finding.

What does the bioweapon report actually say? Anthropic says it blocked misuse attempts over eight months, including a May request to help draft a grant proposal for gain-of-function chikungunya virus research at a military institution. The company could not determine intent, so it erred on the side of caution.

Source: https://www.thepaper.cn/newsDetail_forward_34049099