About 133 million contractor exchanges ran through Claude with bio-risk filters off

The chikungunya request read like a grant application ^[4]. What Anthropic acted on was the setting: the work was to be carried out at a military research institute, and gain-of-function research has legitimate scientific uses ^[5]. Jacob Klein, Anthropic's head of threat intelligence, said, "It's an incredibly nuanced situation." ^[8] He told the New York Times that "You are not seeing someone in a comic book kind of way say, 'Hey, I want to build a biological weapon to kill everybody,'" ^[7].

From May 2025 to April 2026, bio-risk filters were disabled on contractor traffic covering approximately 133 million exchanges with around 50,000 contractors ^[11], which is twelve months ^[4] and about 2,660 exchanges per contractor ^[2]. The retrospective review flagged 1,197 transcripts as high risk, one in roughly 111,000 exchanges ^[12][1]. Anthropic says it found no confirmed evidence of actual bioweapons uplift from that gap ^[12].

The report came out September 10, 2026 and covers roughly eight months of misuse data ^[1]. Counting back, its window opens near January 2026, so it overlaps roughly the last four months of the period when the filters were off ^[3].

The state attributions sit elsewhere in the report. In this account of the report, the blocked biological research goes unattributed ^[5]. Russian state media used Claude to craft propaganda framed as independent reporting, including fabricated claims about a Moldovan election ^[13]. Suspected Chinese and Iranian government-linked actors used the models for dissident surveillance and to target diaspora communities, according to the report ^[14]. Six cases involve software relevant to conventional weapons design, three linked to China, two to Russia and one to Yemen, which the report implicitly ties to the Iran-backed Houthi militia ^[15].

Andrew Weber, a senior fellow at the Council on Strategic Risks and a former U.S. Assistant Secretary of Defense for nuclear, chemical and biological defense programs, reviewed the report before publication ^[16]. He called some findings "chilling examples of state-sponsored biological weapons developers tapping into the rapidly advancing capabilities" of leading AI models, according to the New York Times ^[17]. "The fact that Russia, China and North Korea continue to develop prohibited biological weapons makes it imperative that we deny their researchers access to these extraordinarily capable models," Weber said, according to the Washington Examiner ^[18]. North Korea is not among the states named in the report's misuse cases ^[6]. Susan Monarez, a microbiologist and former acting CDC director, also reviewed the report before publication ^[19].

Regional access controls were already being circumvented. Anthropic says it banned the associated accounts and dismantled the relay networks used to get around those controls ^[3].

Anthropic's own evaluations carry the capability claim: older Claude models could not meaningfully assist in dangerous biological research, and current, more capable models can complete complex scientific tasks ^[10]. The company also says it cannot definitively determine whether blocked research is legitimate dual-use work or weapons development, and that it erred toward blocking ^[9].