Anthropic announced on Thursday that it took action against scientists who utilized its Claude AI models in ways that could potentially facilitate the development of biological weapons. The company released a detailed report outlining these instances, which also covered other types of harmful activity, including those related to scams, conventional weapons development, and public propaganda.
Addressing Biological Misuse
The report disclosed five specific cases involving activity that held the potential to support biological weapons development. Anthropic designated biological misuse as “one of the most serious risks of frontier AI models.” The company emphasized that while the identified incidents provided “evidence of capability” for the AI models, they “cannot concretely demonstrate that such capability would ever be used to develop biological weapons in the real world.”
The individuals involved in these reported activities were working scientists, whose identities Anthropic declined to disclose, stating that it could not “assert that they intended harm.” The firm noted that upon detection and investigation, it banned the users’ accounts and subsequently incorporated its findings into its threat intelligence processes, safeguards, and enforcement mechanisms to prevent and disrupt similar activities in the future.
Anthropic highlighted that newer, more advanced models, such as Claude Fable 5, now feature stronger safeguards that limit access to “a wide range of dual-use biological research queries.” The company stressed the dual nature of such knowledge, explaining that “The same information that can be used to develop a biological weapon could also be used to develop, for example, a vaccine or a cure for a disease.”
Combating Propaganda and Influence Operations
Beyond biological concerns, Anthropic reported identifying and subsequently removing user accounts that were using Claude to conduct influence operations aimed at shaping public opinion. These actions included the removal of three accounts linked to an Iranian state propaganda institution. In these instances, Claude was employed to generate content, make posts appear as if they originated from independent news sources, and spread material across social media platforms such as X, Instagram, and TikTok.
Furthermore, one threat actor connected to Iran utilized Claude to develop recommendations for targeting U.S. naval forces operating in the region. The same account was also responsible for designing software intended for a domestic mass-surveillance system for Iranian state authorities.
Surveillance and State-Sponsored Activities
The report detailed that Claude was also misused by actors in West Africa and China, sometimes for identifying targets or running surveillance operations. In China, three accounts associated with the PRC municipal security service were implicated in activities related to surveillance and transnational repression. Among these was a municipal bureau that specialized in profiling overseas organizations and activists.
Anthropic stated that the AI was sometimes used to replace the need for a dedicated human engineering workforce in these surveillance contexts.
