ع
Anthropic disclosed in its threat intelligence report released September 10, 2026 that it had disrupted five distinct case studies in which users employed its Claude models in ways that could advance biological weapons development, prompting the firm to ban associated accounts and dismantle relay networks used to evade geographic restrictions. The cases, which spanned queries on gain-of-function research involving the chikungunya virus and mammalian adaptation of avian influenza, saw actors obfuscate their purposes and circumvent controls on models including Haiku, Sonnet and Opus, according to the document that drew coverage from AP News and CNN. Anthropic stressed that it could not determine whether the researchers, some apparently linked to state or military institutions, harbored malicious intent or pursued legitimate scientific aims, yet chose to act because the potential consequences demanded caution.
Biological misuse ranks among the most serious risks of frontier AI systems, the company stated in the report, noting that without adequate safeguards such capabilities could have catastrophic consequences while the same underlying information might support vaccine development or disease cures. “You are not seeing someone in a comic book kind of way say, ‘Hey, I want to build a biological weapon to kill everybody,'” Jacob Klein, head of threat intelligence at Anthropic, told The New York Times, underscoring the nuanced overlap between beneficial inquiry and dual-use technologies. Over a sampled 30-day window the firm identified roughly 35 distinct research efforts displaying concerning patterns, though most did not involve its most powerful models such as Claude Fable or Mythos-class systems.
The report, Anthropic’s third on AI misuse since March 2025, also catalogued disruptions of Russia-linked cyber espionage campaigns and Iranian propaganda operations that harnessed Claude for content creation and influence activities. In parallel the company tracked attempts to use the models for conventional weapons design including missiles, armed drones and targeting systems as well as surveillance, scams and unauthorized model distillation. A separate earlier Anthropic assessment had revealed that biological risk filters remained disabled on contractor platforms for approximately 133 million exchanges between May 2025 and April 2026, a gap that prompted a retrospective review flagging more than 1,100 high-risk transcripts.
To counter evolving threats Anthropic has rolled out stronger safeguards in recent models including Claude Fable 5, which restrict access to a wider range of dual-use biological queries while an August 2026 classifier update reduced unnecessary fallbacks by about 85 percent across its platforms. The firm shared investigative findings with other AI laboratories and government authorities to bolster collective defenses, emphasizing that as models grow more capable even individual actors can generate risks once requiring institutional resources. Coverage from Forbes and USA Today highlighted that the blocked grant-application query for chikungunya research appeared tied to a military-linked laboratory, illustrating how seemingly routine scientific requests can signal broader clusters of activity.
Anthropic’s actions reflect an industry-wide push to refine safety layers through constitutional classifiers, red-teaming and collaboration with external experts on chemical, biological, radiological and nuclear risks. Experts cited in AP News reporting have called for expanded government regulation rather than sole reliance on voluntary corporate measures as AI capabilities continue to advance. The company indicated that evidence suggests its existing protections are functioning yet acknowledged that more robust systems will be required to match the sophistication of future models and emerging threat vectors.
