Anthropic released a threat‑intelligence report on Thursday that lists over four decades of misuse of its Claude models for malicious activity that could support the development of biological weapons. The firm says it has identified and disrupted attempts to use AI for cyber‑crime, political influence, scams, surveillance and to design steps for synthesising weaponised biological agents.
The report highlights five high‑profile instances where actors used Claude to produce technical guidance for creating harmful biological weapons and six occasions where the system was used to develop or refine conventional weaponry, such as firearms, missiles and armed drones.
The assessment is part of a growing wave of warnings from industry and safety researchers. That wave includes the US Senator Bernie Sanders’ call to pause advanced AI development and a bipartisan push for an international treaty on super‑intelligence safety, as well as OpenAI chief scientist Jakub Pachocki’s appeal for voluntary slowdowns until robust safeguards are in place.
In its analysis, Anthropic points out that biological weapon research remains the most serious danger posed by frontier AI — because the same information that could be used to develop a virus could also help scientists create vaccines and cures. It stresses that without the right controls, the threat spectrum could expand to include cyber‑security breaches, disinformation campaigns and finance‑fraud attacks, many of which hinge on AI automation and learning.
The company shared its findings with law‑enforcement agencies and other AI firms, and has upgraded its monitoring protocols to better detect and block future malicious activity. “It’s an incredibly nuanced situation,” said Jacob Klein, Anthropic’s head of threat intelligence, explaining that the model can “preserve a dangerous knowledge base” even if the user travels back to the intent of doing harm.
This push for stronger governance comes as tech giants such as Google and Meta have also begun publishing their own AI misuse reports. Google warned that a private user tried to obtain a step‑by‑step guide for producing lethal biological agents via Gemini, and Meta’s AI platform contains safeguards to block harmful instructions in real‑time. Critics argue however that reactive damage control is too late; they demand clear international rules and a global AI safety mandate that go beyond voluntary self‑regulation.
