Days After ‘AI Could Kill Us All’ Warning, Anthropic Reveals Bioweapon Threat as Altman Says OpenAI Could Slow Down

Just days after former OpenAI and Anthropic researcher Jacob Coxon stunned the technology industry by warning that artificial intelligence could «kill us all by the end of the decade,» developments inside two of the world's most powerful AI companies are putting a renewed spotlight on the dangers their own researchers have been describing. Anthropic has disclosed troubling evidence involving attempts to use advanced AI in connection with potentially dangerous biological work, while OpenAI CEO Sam Altman has told employees his company could be willing to slow development of cutting-edge systems under certain conditions. The developments do not constitute a coordinated response to Coxon's resignation, but their timing is striking. Coxon had accused both companies of continuing an increasingly dangerous race despite understanding the stakes, declaring that they are «racing straight to self-improving superintelligence and gambling with our lives.» Now, the companies themselves are confronting concrete questions about how increasingly capable AI could be misused and whether the race should continue at its current speed.

Anthropic's latest findings provide a particularly disturbing example of why AI safety researchers have become increasingly focused on biological threats. The company detected activity involving its systems that raised concerns about users seeking assistance potentially relevant to biological weapons, forcing Anthropic to intervene and block the activity. The company could not definitively determine whether the individuals involved had malicious intentions, an important distinction that prevents the episodes from being characterized as confirmed attempts to manufacture bioweapons. Nevertheless, the cases demonstrate a problem that becomes more serious as models acquire greater scientific capabilities: information that can accelerate legitimate biological research can potentially also assist someone pursuing dangerous objectives. That possibility closely resembles the broader category of catastrophic misuse that AI executives themselves have warned about for years. Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis previously joined other prominent researchers in declaring that «mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.» Anthropic's new disclosure makes the biological component of that warning considerably less abstract.

The disclosures also arrive as OpenAI appears increasingly willing to publicly contemplate something that Coxon specifically argued the industry may eventually need: slowing the race. Altman has told OpenAI employees that the company is open to coordinating with competing frontier laboratories to reduce the pace of cutting-edge AI development if safety conditions warrant it, a notable position from the chief executive of one of the companies pushing hardest toward increasingly capable systems. The idea does not amount to OpenAI announcing a unilateral halt, nor does it mean the broader competitive race has ended. But it directly touches the contradiction at the center of Coxon's criticism: companies acknowledge potentially catastrophic risks while simultaneously fearing that slowing independently could allow a competitor to move ahead. That concern was reinforced by Anthropic Alignment Science Lead Evan Hubinger, who publicly supported Coxon's warning. «Jacob is correct here — we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade,» Hubinger wrote, while acknowledging that Anthropic still lacks a reliable solution for aligning a future superintelligence with humanity's interests.

he third development adds another layer to that paradox. Altman is pitching American utility companies on using OpenAI's technology to strengthen the electrical grid against increasingly sophisticated AI-enabled cyberattacks, effectively presenting advanced artificial intelligence as part of the defense against threats that more capable AI systems could themselves intensify. The proposal reflects growing concern that critical infrastructure — including electricity networks, communications systems and other essential services — could become increasingly vulnerable as AI lowers the technical barriers for sophisticated cyber operations. OpenAI and Anthropic have already warned that rapidly improving models are becoming significantly more capable in cybersecurity, while recent incidents have demonstrated that autonomous agents can sometimes circumvent controls and behave in unexpected ways. The result is an unusual technological arms race: AI may provide defenders with powerful new capabilities for identifying and responding to attacks, while simultaneously giving adversaries tools capable of finding vulnerabilities, automating operations and attacking infrastructure at unprecedented speed.

The warnings are particularly striking because they are no longer coming exclusively from outside critics of the AI industry. Executives and researchers responsible for developing frontier models have repeatedly acknowledged scenarios in which sufficiently advanced systems could produce catastrophic consequences. Altman, Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis were among the prominent figures who signed the declaration that «mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.» Amodei has separately warned about powerful networks of AI systems potentially becoming capable of taking control on an extraordinary scale, while OpenAI Chief Scientist Jakub Pachocki has argued that «intervention may be needed to ensure humans remain in control of the future.» Even Elon Musk has publicly estimated substantial odds of catastrophic AI outcomes. The remarkable feature of the current debate, therefore, is not whether influential technology leaders acknowledge serious risks; it is how they reconcile those warnings with the continuing race to build substantially more capable systems.

That contradiction is precisely what made Coxon's resignation so provocative. His argument was not simply that advanced AI might eventually become dangerous, but that people inside the companies developing it already understand the potential consequences and continue racing forward because each laboratory fears slowing down while its competitors advance. UC Berkeley computer science professor Stuart Russell has described that dynamic in even harsher terms, declaring that «for governments to allow private entities to essentially play Russian roulette with every human being on earth is, in my view, a total dereliction of duty.» Against that backdrop, Anthropic's intervention against potentially dangerous biological activity, OpenAI's push to protect critical infrastructure and Altman's willingness to contemplate coordinating a slowdown take on greater significance. None proves that catastrophe is approaching, and none represents an abandonment of frontier AI development. But together they sharpen the question Coxon placed before the industry: if the companies building the technology increasingly acknowledge that the risks could become catastrophic, how far are they actually prepared to go to prevent them?

Created by humans, assisted by AI.