Anthropic caught scientists probing bioweapons with Claude

Five case studies show how Anthropic spotted working scientists pushing its models toward dangerous biology, and what it did next.

By ·

Sep 10, 2026

6 cards · 1m 30s · tap to begin

From · · · 1m 30s

Anthropic caught scientists probing bioweapons with Claude

Five case studies show how Anthropic spotted working scientists pushing its models toward dangerous biology, and what it did next.

In brief

Five case studies show how Anthropic spotted working scientists pushing its models toward dangerous biology, and what it did next. The hard part is not catching obvious villains, it is telling vaccine research apart from bioweapon work, so Anthropic now errs toward over-blocking. Originally reported by Engadget.

Caught in the act

The App Store page for Claude, Anthropic

Anthropic says it flagged and cut off multiple scientists who were using its models to work toward potential biological weapons. The detail landed in a misuse report published Thursday.

Why now

Biological misuse is one category among many in the report, but it matters more each release. Newer models like Fable 5 are getting genuinely good at real scientific research.

Vaccines look like weapons

Detection is hard because the two lines of work can read almost identically in a prompt. Anthropic told the New York Times it chose to be overly cautious every time, given what is at stake.

“You are not seeing someone in a comic book kind of way say, 'Hey, I want to build a biological weapon to kill everybody.'”

The chikungunya grant

The chikungunya grant

In May, a safety classifier caught a request to write a grant for gain-of-function work on chikungunya, a virus with no licensed treatment. The proposal aimed to boost transmissibility and immune evasion.

Bird flu and venom peptides

Two other cases: more gain-of-function work on bird flu, and a researcher building an atlas of venom toxin peptides plus a pipeline to optimize toxin traits.

“a generative pipeline that optimized toxin characteristics”

Banned but not named

Every implicated account was banned, and the findings now feed new safeguards. Anthropic withheld names because these are working scientists it does not accuse of intending harm.

“identifying them or their labs could expose them to harm”

The key point

The hard part is not catching obvious villains, it is telling vaccine research apart from bioweapon work, so Anthropic now errs toward over-blocking.

Read the original on Engadget

React

Sign in to react and comment.

Comments (0)

Life is short. Keep it sweet. Respect others' opinions and be kind!

    Recommended next

    More decks on ai safety and related topics.