Anthropic says it blocked five separate attempts this year to use Claude for bioweapons research.
The company published five case examples in which users tried to get around Claude's safety controls, using tactics it describes as efforts to obfuscate the true purpose of their research. Some of the users were based in countries Anthropic bars from accessing its models, including Russia, China, and Iran. Anthropic did not detail exactly what each project entailed or how far the research got before being shut down. The disclosures appeared in a broader report the company published on misuse of its models.
Biological weapons research is one of the scenarios AI safety researchers worry about most, since a capable model could in theory walk someone without specialized training through steps that used to require a well-funded lab. That five such attempts surfaced in a single year, despite geographic access restrictions already in place, suggests those restrictions are more speed bump than wall. It also shows Anthropic's safeguards are being actively probed by people motivated enough to disguise their intent, not just tripped by casual misuse.
Anthropic says the goal of publishing the examples is to spark a conversation among AI companies and governments about biological risk. The report does not say what concrete countermeasures, if any, follow from that conversation.