Anthropic, the AI startup behind the Claude family of language models, announced Thursday that it has intercepted multiple attempts by bad actors to exploit its technology for dangerous biological research. The company said the incidents, which span from December 2025 through August 2026, involved requests to use its older Claude Opus 4 and Claude Sonnet 4.5 models to draft grant applications for gain‑of‑function studies on the chikungunya virus.
Misuse cases and the company’s response
In the most notable case, a user asked Claude to help write a scientific funding proposal that would enhance the transmissibility and immune‑evasion properties of chikungunya, a mosquito‑borne virus that already causes severe pain and fever. Anthropic’s systems flagged the request and blocked it, noting that while such research could lead to better vaccines, it also carries the risk of creating a more dangerous pathogen.
The company’s third misuse report since March 2025 also identified other actors—including spyware vendors, politically motivated individuals, and state‑sponsored groups—who tried to harness its AI for cyber‑attacks, surveillance, and propaganda campaigns. Nine coordinated influence‑operation cases were traced to networks operating out of Russia, Iran, Turkey, the Persian Gulf, South Asia, Africa, and Europe.
New safeguards for newer models
Anthropic emphasized that none of the reported incidents involved its newest, more powerful Claude Fable or Mythos‑class models, except for a single “industrial‑scale, covert campaign” that attempted to extract model capabilities without authorization. Because the newer models can assist with complex scientific tasks, the company has now applied stronger safeguards that restrict a broad range of dual‑use biological queries.
According to the report, older models were “well below the threshold where they could meaningfully assist a sophisticated user in carrying out dangerous biological research,” so their safeguards focused mainly on preventing novice attempts to recreate known bioweapons. The updated protections aim to block both novice and advanced misuse attempts.
Industry and academic reactions
Experts have long warned that self‑regulation may be insufficient. John Thickstun, an assistant professor of computer science at Cornell University, called the situation “uncomfortable” for companies like Anthropic and OpenAI, noting the lack of democratic oversight in deciding what constitutes safe versus unsafe AI behavior.
Anthropic’s report also highlighted a recent resignation by researcher Jacob Coxon, who expressed concerns that the AI race toward self‑improving superintelligence could endanger human life by the decade’s end. Coxon’s departure underscores internal tensions about the pace of AI development and the need for responsible safeguards.
Government and industry collaboration
Anthropic said it has shared its findings with government authorities and industry partners, urging them to recognize similar patterns on their platforms. The company hopes the report will help other developers strengthen defenses and give policymakers a clearer view of emerging AI threats.
As AI models become more capable, Anthropic warned that the risk landscape will expand unless developers, regulators, and civil‑society defenders act together to make the technology safer.
Original reporting: NBC Connecticut — read the source article.
