Anthropic reports blocking of AI accounts linked to potential bioweapon research
Anthropic said it disabled several user accounts that attempted to use its Claude models for work that could aid the development of biological weapons, citing five case studies uncovered over a month.
Anthropic disclosed that it had terminated multiple user accounts after detecting attempts to employ its Claude artificial-intelligence models for research that could support the creation of biological weapons. The company’s Thursday report described five real-world cases in which users bypassed geographic controls and concealed the purpose of their queries to evade safeguards. The incidents involved work on gain-of-function studies of viruses such as chikungunya and bird flu, as well as research on orthopoxviruses and venom toxins.
Reviewing roughly 30 days of activity, Anthropic identified about 35 distinct research efforts that raised concerns, though it could not confirm whether the actors intended malicious outcomes or were conducting legitimate scientific work. The report noted that “sophisticated attacks no longer require sophisticated attackers,” highlighting how AI models lower the barrier between well-resourced state actors and individual operators.
Anthropic said its newer models, including Claude Fable 5, incorporate stronger safeguards that limit access to dual-use biological queries. While the company believes the safeguards have been effective so far, it warned that more robust safety systems will be needed as AI capabilities and potential threats evolve.
“If you extrapolate into the future, the level of capabilities of these AIs … they could cause extreme havoc,” former Anthropic employee Jacob Coxon told CNN.
The disclosure comes amid broader industry concerns about AI safety, with former employees and researchers urging tighter regulation and slower development to mitigate risks of misuse in areas ranging from bioweapons to surveillance and conventional weapons.