Anthropic discussed attempts to use AI to develop biological weapons

Anthropic, the world’s most valuable artificial intelligence startup, reported that this year it identified and blocked several attempts to use its models for research that could potentially aid in the development of biological weapons. The company described five such cases, warning of growing risks as AI models become more powerful.

In these cases, users attempted to circumvent the restrictions built into the AI models and conceal the true purposes of their research, Anthropic stated. The company published a report titled “Detecting and Combating AI Abuse” for September 2026. Some of the users were located in countries where the company prohibits access to its models, including Russia, China, and Iran.

In one of the most serious cases, a researcher from an “unsupported region” spent several weeks using Claude’s AI to plan experiments with the avian flu virus. Anthropic’s security system restricted his access, leaving him with only the weakest models.

The company emphasized, however, that it cannot assert that the scientists involved in the cited cases actually intended to cause harm. According to the company, the same knowledge and tools that can be used to create biological weapons are also applicable to vaccine development. Anthropic suspended the accounts mentioned in the report but did not disclose the names of the research organizations or the countries where the incidents occurred.

"We hope that by sharing these examples, we can spark a discussion within the AI industry and among governments about emerging biological risks and ways to address them," the company said.

The debate over AI safety intensified earlier this week: On September 9, former Anthropic employee Jacob Coxon stated after leaving the company that AI development teams “seriously believe it could kill us all by the end of the decade.” Anthropic researcher Evan Hubinger agreed with him: he wrote on the social media platform X that, in his estimation, “the probability of such a scenario is more than 10% over the next decade.”

Concerns about the technology began to grow earlier this year following the development of advanced models such as Anthropic’s Mythos, according to the Financial Times. In July, OpenAI also caused alarm in the market by reporting that its models had autonomously hacked the AI startup Hugging Face.

Executives at AI companies and biosafety experts are increasingly speaking out about the need to regulate the use of such technologies, as they could potentially be exploited by terrorist organizations, government agencies, or even lone actors to develop biological weapons or spread dangerous pathogens, the FT reports.

However, the newspaper notes that there remain significant technological and resource-related barriers between the creation of a theoretical design for a biological weapon using AI and its actual production.

Anthropic also reported other instances of misuse of its technologies—ranging from a network of fake dating apps designed for fraud to systems used to monitor dissidents. In addition, on September 10, the company stated that seven Chinese labs, including Moonshot and DeepSeek, had attempted to replicate the capabilities of its models using a technique known as “distillation” — a method in which a developer uses data from a more powerful and larger AI model to improve its own less powerful system.

This article was AI-translated and verified by a human editor