‘AI Could Kill Us All’ Warning Gains New Weight as Anthropic Flags Bioweapon Threat and Altman Talks Slowdown

Days after former OpenAI and Anthropic researcher Jacob Coxon warned that artificial intelligence could «kill us all by the end of the decade,» developments at two of the world’s most powerful AI companies are putting fresh attention on the dangers their own researchers and executives have repeatedly acknowledged while continuing to develop increasingly capable systems.

Coxon’s resignation was particularly striking because he had worked inside both OpenAI and Anthropic. Rather than accusing the industry of simply overlooking potential dangers, he argued that leading laboratories understand them and continue competing anyway, accusing the companies of «racing straight to self-improving superintelligence and gambling with our lives.»

Anthropic has now disclosed activity involving its AI systems that raised concerns about assistance potentially relevant to dangerous biological work. The company intervened and blocked the activity, although it could not determine conclusively whether those involved had legitimate scientific purposes or malicious intentions, an important distinction when assessing the seriousness of the incidents.

The episodes illustrate a difficult problem facing increasingly sophisticated AI models. Systems capable of helping scientists accelerate legitimate biological research could potentially provide useful information to someone pursuing harmful objectives. As AI becomes more capable, companies must therefore determine when ordinary scientific assistance crosses a line into knowledge that could facilitate biological weapons or other dangerous activities.

The concern is hardly limited to outside critics. OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis signed an industry declaration stating that «mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war,» alongside numerous other leading researchers and executives.

Against that backdrop, Altman has reportedly told OpenAI employees that the company could be willing to coordinate with rival frontier laboratories to slow cutting-edge AI development if circumstances require it. Such a move would not represent an end to the AI race, but it signals that even its most influential participants are contemplating limits on its speed.

Anthropic Alignment Science Lead Evan Hubinger publicly reinforced Coxon’s extraordinary 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, underscoring how fears of catastrophic outcomes exist among researchers working directly on advanced AI safety.

Altman is simultaneously pitching American utilities on using OpenAI technology to help defend the electrical grid against increasingly sophisticated AI-enabled cyberattacks. The proposal creates a striking paradox: advanced AI could become an important defensive weapon against cyber threats even as increasingly capable models potentially give attackers powerful new tools for targeting critical infrastructure.

That dynamic could produce an escalating technological contest between attackers and defenders. AI may help utilities identify vulnerabilities and respond rapidly to cyberattacks, while similar technology could potentially allow adversaries to automate operations, discover weaknesses and increase their capabilities. The same technological advances could therefore strengthen both sides of an increasingly consequential cybersecurity battle.

The debate is especially unusual because warnings are coming from people building the technology. OpenAI Chief Scientist Jakub Pachocki has said that «intervention may be needed to ensure humans remain in control of the future.» Meanwhile, executives overseeing frontier laboratories continue investing enormous resources in developing systems designed to become substantially more capable than today’s models.

Anthropic blocking potentially dangerous biological activity, OpenAI proposing AI defenses for critical infrastructure and Altman contemplating a coordinated slowdown do not mean the industry has abandoned its race toward more powerful AI. Instead, they sharpen Coxon’s central challenge: if the people developing advanced AI genuinely believe the consequences could become catastrophic, how far are they actually prepared to go to prevent them?