Anthropic AI Risk Warnings Raise Questions on Safety and IPO Disclosures
AI researcher Jacob Coxon resigned from Anthropic, warning that leading AI companies were “gambling with our lives.” The departure intensified after Anthropic’s alignment lead publicly stated that AI could kill all humans and put the likelihood at more than 10% within the next decade.
In an episode of TechCrunch’s Equity podcast, Kirsten Korosec, Sean O’Kane and Anthony Ha examined the renewed debate over existential AI risk. Their discussion considered the meaning of such public warnings, the limits of numerical claims about catastrophe, and the potential implications as Anthropic moves towards a possible initial public offering.
A warning that broadened the AI safety debate
Coxon’s decision to leave distinguished his intervention from warnings issued by executives or researchers who remain employed by companies building increasingly capable models. Ha said that, if Coxon genuinely believes the work poses an extreme threat, resigning places his professional trajectory behind that conviction.
At the same time, Ha challenged the usefulness of assigning a precise probability to an outcome such as human extinction. He described the greater-than-10% figure as unsupported, while noting that it may refer to the “P(doom)” framing used in parts of the AI safety community.
The discussion followed reports and public debate around AI agents, as well as recent releases from Anthropic and OpenAI. O’Kane argued that some episodes can leave the impression that companies do not fully control their systems, rather than presenting a polished narrative solely designed to demonstrate technical capability.
Capability messaging and investor disclosure
Korosec raised a different possibility: statements about dangerous or unexpectedly capable systems can also function as a signal of technological advancement. In that reading, warnings that a model can break through limitations may communicate strength and sophistication, even when they describe a potential failure.
Ha did not dismiss genuine concern among researchers and executives, but said that apocalyptic language can also align with commercial interests. A company that presents its software as exceptionally powerful may reinforce the perception that its technology is important, even if the same description raises questions about safety.
The timing matters because Anthropic was discussed as being weeks or months from a potential IPO. O’Kane questioned how existential-risk claims may appear in an S-1 registration statement, particularly in the risk-factor language investors use to understand material business exposures. Korosec noted that, in the current investment environment, language suggesting danger may not necessarily reduce a company’s valuation.
Immediate harms remain part of the picture
Ha argued that debate over AGI and superintelligence can consume attention that should also be directed at nearer-term harms. He identified labour, environmental and climate-related effects as risks that require discussion alongside longer-horizon concerns about advanced AI systems.
The practical business implication is to separate public claims about hypothetical existential probabilities from risks that can be identified, governed and disclosed today. Organisations deploying AI should examine the controls around their systems and address operational, workforce and environmental impacts without allowing broad catastrophe narratives to displace concrete safeguards.