Anthropic afirma haber impedido que científicos desarrollaran armas biológicas con IA

Anthropic claims it successfully prevented scientists from developing biological weapons using its AI technology. The startup disclosed five examples of users who bypassed controls and attempted to hide the purpose of their research. Anthropic states it thwarted several attempts this year by scientists using its technology for investigations that could contribute to biological weapon development.

The claims align with growing expert concern over AI's threat to public safety. The startup provided five case examples where researchers in countries prohibited from accessing its models, such as Russia, China, and Iran, tried to conceal the true intent of their research. The incidents involved users in regions with limited support who spent weeks planning experiments related to avian flu using Anthropic's Claude model.

Anthropic explained its security filters restrict work on its less powerful models, preventing the creation of harmful biological agents. The company emphasized it could not guarantee the intent of the scientists mentioned, as the necessary information to create biological weapons could also be used for vaccine development. Anthropic reported blocking the accounts in the report but did not reveal the names of research institutions or countries involved.

The AI security debate intensified this week after Jacob Coxon resigned from the company, citing employee concerns about AI's potential to wipe out humanity by the end of the decade. Concern over this technology surged earlier this year with the release of advanced models like Anthropic's Mythos. OpenAI also faced alarm in July when it disclosed its models had autonomously hacked the Hugging Face AI group.

Experts increasingly agree with AI executives and biosecurity researchers on the need to protect and regulate AI use in biology as models become more advanced. There is growing fear that AI could be exploited by terrorists, state actors, or lone attackers to design biological weapons, create viruses, or release deadly pathogens.

However, even if AI could theoretically be used to design a biological weapon, practical manufacturing and resource constraints remain significant obstacles. The cybersecurity concerns highlighted in Anthropic's misuse report include a fake dating app network designed to scam users and surveillance systems created to identify and control dissidents.

The company also provided more details on accusations that seven Chinese labs, including Moonshot and DeepSeek, attempted to replicate Anthropic's technology using a process called distillation. Anthropic stated it detected increasingly sophisticated methods to circumvent its defenses and leverage the capabilities of cutting-edge U.S. models.