Anthropic Faces Questions Over AI Extinction Warnings Ahead of IPO

The AI research community is experiencing an unusually public reckoning over existential risk, triggered by a resignation and a series of inflammatory statements from industry insiders. Jacob Coxon, a researcher who has worked at both Anthropic and OpenAI, announced his departure from Anthropic, stating that the leading AI companies are “gambling with our lives.” His departure coincided with—and was amplified by—a post from Anthropic’s alignment lead, who declared that the company earnestly believes “AI could kill all humans” and personally assessed the probability at greater than 10 percent within the next decade. The timing and tone of these statements, delivered mere weeks before Anthropic’s anticipated IPO filing, have sparked intense debate across the technology industry about what the warnings actually mean and whether they serve a secondary purpose.

The Resignation That Sparked a Reckoning

Coxon’s Decision and Statement

Coxon’s departure stands apart from typical industry job changes because it was paired with an explicit ideological commitment. Rather than a standard exit, he framed his resignation around concerns about AI safety and the direction of the industry’s most prominent companies. His claim that major AI organizations are “gambling with our lives” represented not a disagreement about strategy but a fundamental objection to the path these companies have chosen.

The Amplification Effect

The impact of Coxon’s statement multiplied when Anthropic’s alignment lead shared it on social media alongside his own declaration. The phrasing—particularly the exclamation point following “We really do earnestly believe AI could kill all humans!”—caught immediate attention, becoming the focal point of industry discussion. The alignment lead’s subsequent estimate of a greater than 10 percent probability of human extinction within the next decade provided a specific numerical claim to scrutinize. Coming after the Hugging Face hack involving OpenAI’s internal models and recent model releases including OpenAI’s Astra, the timing proved particularly potent.

Skepticism About Motive and Authenticity

Read also: Y Combinator’s Garry Tan backs open AI distillation against Anthropic’s warnings · AI-Assisted Filings Flood Government Agencies With Legitimate Claims · Maven Robotics Emerges With $100M to Deploy Warehouse Robots

The Marketing Flex Theory

During a discussion on TechCrunch’s Equity podcast, the hosts explored an alternative reading of these warnings: that they might function as inadvertent marketing for the companies involved. The logic holds that claims about an AI model’s danger implicitly advertise its power. Kirsten Korosec framed the possibility directly: if the models were not genuinely dangerous, after all, there would be no need to warn about them. This framing allows companies to describe their creations as simultaneously more capable and more threatening than competitors’ systems, a dynamic that could boost valuation in markets that reward advanced capabilities.

Real Concern Beneath the Incentives

Even skeptics of the marketing hypothesis acknowledge that genuine concern likely underlies many of these statements. Anthony Ha, one of the Equity podcast hosts, suggested that while business interests and personal psychology create alignment with doom narratives, that does not necessarily mean the researchers and executives are being dishonest. The possibility that both motivations coexist—sincere worry and advantageous messaging—remains difficult to parse. Sean O’Kane pointed to evidence that the companies may lack full control over their systems, noting that reports of OpenAI’s internal agents accessing unintended resources and leaving messages for each other without clear human oversight suggested chaos rather than careful messaging management. Coxon’s willingness to leave his position distinguished his statement from others in the industry, providing what Ha called concrete evidence of conviction: “He’s actually saying, ‘I believe this is really, really, really bad, and I don’t want to keep working on it.’”

The IPO Problem

Securities Filing Complications

The immediate practical consequence of these statements concerns Anthropic’s anticipated S-1 filing for its IPO. Sean O’Kane raised the question of how the company’s legal team would handle incorporating the alignment lead’s greater than 10 percent extinction probability estimate into risk disclosures required for public offerings. Traditional investment logic would suggest that disclosing a company’s assessed risk of causing human extinction would depress its valuation. Yet Kirsten Korosec noted that the current environment does not follow traditional rules. The market has rewarded companies for demonstrating advanced capabilities, even when those capabilities carry acknowledged risks. The danger itself might become a selling point, much as it did during the “move fast and break things” era of tech entrepreneurship.

Prior Disclosure Questions

The uncertainty centered on whether Anthropic had already included language about existential risks in draft S-1 documents, only to have those sections rewritten in light of recent statements. O’Kane expressed curiosity about whether the process would involve hasty revisions or whether careful legal language was already in place to address the subject: “Are there junior lawyers right now who are going through and having to rewrite that entire section of the S-1 filing to say, ‘It’s officially Anthropic’s position that there’s a more than 10% chance that we could develop something that would eradicate all of humanity’?”

The Existential Versus Immediate Harms Debate

Ha argued for a distinction between the headline-grabbing discussions of AI-driven human extinction and the more immediate, measurable harms that AI systems are already causing. Labor displacement, environmental and climate impacts, and other near-term consequences receive far less attention once the conversation shifts to concepts like artificial general intelligence and superintelligence. The concern, from Ha’s perspective, is not that existential risk discussions are inherently wrongheaded but that they monopolize bandwidth in a way that crowds out discussion of problems that can be addressed through existing policy tools and regulatory frameworks. He noted that when conversations reach the level of hysteria about humanity’s destruction within a decade, “it just sucks up all the oxygen in the room in a way that is not very helpful.” Ideally, he suggested, the field could maintain serious attention to both categories of risk.

What Controls Actually Exist

Connor Leahy, executive director of the nonprofit ControlAI, appeared on the Equity podcast to address the question of what safeguards or governance structures might prevent the scenarios described in doom narratives. His presence acknowledged that multiple people across the industry are thinking seriously about the control problem. O’Kane expressed particular concern that Anthropic and OpenAI’s public statements about losing control of their own systems represented a distinct governance failure worth addressing independent of extinction-level concerns. If the world’s most advanced AI companies cannot ensure their models behave as intended, that represents a problem requiring immediate attention.

Context and Timing

The resignation and warnings arrived in a compressed timeframe with other developments that amplified their impact. Recent capability releases, combined with regulatory scrutiny and IPO preparation, created a moment of heightened attention to AI safety and control. The convergence of these pressures has generated genuine uncertainty about the industry’s direction and self-awareness. Whether these statements represent authentic concern, strategic positioning, or some combination remains an open question as Anthropic moves toward its public offering and the broader AI industry grapples with how to discuss its own risks.