Anthropic Blocks Attempts to Bypass Claude's Safeguards for Bioweapons Research

Anthropic has reported that this year it repeatedly blocked attempts to use its technology for research that could help create biological weapons. In a report on the misuse of its models, the company gave five examples in which users “circumvented controls” and tried to “disguise” the true purpose of their work in order to trick its safeguards. Some of these users were in countries barred from accessing Anthropic's models — among them Russia, China and Iran.

“We hope that by sharing these examples we will spur a conversation within the industry and with governments about novel biological risks and how to counter them,” the report says.

One of the cases described involved a researcher from an “unsupported region” who spent “weeks planning” experiments related to bird flu with Claude's help. According to the company, its safety filters limited this work to only its weakest models. All of the accounts mentioned were blocked, but Anthropic did not disclose the names of the institutions or countries involved.

Where the line runs

The main difficulty, the company acknowledges, is that it cannot be certain of the malicious intent of those involved: the very information needed to create a biological weapon is equally suited to developing a vaccine. And even if an AI is capable of designing a theoretical weapon, practical barriers remain — being able to manufacture it and find the resources to do so.

The report comes amid intensifying debate over AI safety. This week Jacob Coxon resigned from Anthropic, stating that employees “genuinely believe AI could kill us all by the end of the decade.” Concern had already grown earlier this year with the release of advanced models, and in July OpenAI caused alarm by reporting that its models had autonomously hacked the Hugging Face platform.

Not just biology

In the same report, Anthropic listed other abuses as well: a network of fake dating apps built to defraud users, and surveillance systems designed to identify and monitor dissidents.

Separately, the company provided new details on how seven labs from China, including Moonshot and DeepSeek, had tried to replicate its technology through distillation. Anthropic said it was seeing “increasingly sophisticated ways of getting around our safeguards and adopting the capabilities of frontier American models.”

Among biologists and AI-company executives, a consensus is growing: as models advance, the use of AI in biology must be protected and regulated. Experts fear that the technology could be exploited by terrorists, state actors or lone individuals to create viruses or to release existing dangerous pathogens.

Source: Ars Technica