Anthropic has reported that it discovered and disrupted several attempts to use its Claude AI models for building biological weapons.
In a briefing released Thursday, the company detailed how attackers had sought step‑by‑step instructions for creating weaponised agents, and how its systems were used by state‑backed groups, cybercriminals and even propaganda outfits linked to Russia and Iran.
Those incidents all occurred between December 2025 and August 2026, and the firm says none involved its new Mythos class of models, which are most powerful. Only one instance of distillation—training smaller models from larger ones—was detected.
“Biological misuse is one of the most serious risks of frontier AI,” said chief threat‑intel officer Jacob Klein. “Without safeguards, this tech could enable catastrophic consequences.”
The report also recounts successful prevention of misuse for conventional weapons design, surveillance, scams and fraud, and says the firm has blocked scientists whose work could support the creation of deadly strains.
Anthropic has shared its findings with authorities and industry partners and stresses that other big players—Google, OpenAI, and others—are drafting similar threat‑intelligence updates to demonstrate their mitigation efforts.
Across the tech sphere, leaders are demanding broader intervention: Senator Bernie Sanders pushes for a pause on super‑intelligence, and the Trump administration worries that dropping behind in AI could destabilise national security.
Anthropic’s move underscores a growing industry trend to provide transparent threat updates and urges governments to develop a safety treaty on AI development and deployment.
