Anthropic flagged scientists using Claude on dual-use biological research and shut the accounts

Anthropic published a September 2026 threat intelligence report covering operations it identified and disrupted from December 2025 through August 2026, the company writes in Detecting and countering misuse of AI. Biological misuse is one of seven harm areas in the file.

Ian Carlos Campbell at Engadget writes that Anthropic flagged and stopped multiple scientists using Claude on work that could further biological weapons research. The report includes five biological-misuse case studies.

The models in these cases were Claude Haiku, Sonnet, and Opus, the widely available tiers. Anthropic says none of the misuse cases in the broader report involved Claude Fable or Mythos-class models, with the exception of one illicit distillation case elsewhere in the file.

Jacob Klein, Anthropic's head of threat intelligence, told the New York Times via Engadget that the pattern is nuanced dual-use work, not comic-book language about building a weapon to kill everybody.

Vaccine research can look like bioweapon research. Anthropic told the Times it erred toward caution.

Gain-of-function means lab work that makes a pathogen more transmissible or harder for the immune system to stop. That work can support vaccine research. It can also raise weapons concern.

Dual-use means the same research trail can help or harm depending on who is using it and why.

Engadget notes from the Anthropic report include a May classifier flag on a grant draft for gain-of-function research on chikungunya, a disease with no licensed treatment, including proposals that increased transmissibility and immune evasion and an affiliation with a military research institute. Similar concern appears on bird-flu gain-of-function. A separate case covers a venom toxin peptide atlas and a generative pipeline optimizing toxin characteristics.

Accounts were banned. Individuals and labs were not named. Anthropic's line: we do not assert that they intended harm.

Ina Fried at Axios covers other chapters of the same report, including state-linked surveillance operations. Those are separate mechanisms from the biological cases.

The disclosure buys a public dual-use record without a proven-intent verdict. Frontier labs can shut the account when the classifier cannot tell the difference. They still cannot put a motive on the unnamed scientist from outside the chat log.

Caution closed the accounts. Intent stayed unproven.

—

Sources

• Anthropic – September 2026 threat intelligence report
• Engadget – scientists / biological weapons research cases
• Axios – same report / surveillance chapter