AI safety firm Anthropic has reported that it intercepted and blocked several attempts to use its Claude models for the development of biological weapons, describing such misuse as one of the most significant risks posed by frontier artificial intelligence.
In its September 2026 threat intelligence report, released Thursday, the California-based company detailed disruptions to “malicious activity” involving its Claude Haiku, Sonnet, and Opus models between December 2025 and August 2026. The incidents included cases where scientists or actors sought information that could facilitate biological weapon creation.
Jacob Klein, Anthropic’s head of threat intelligence, told the New York Times that the situation is “incredibly nuanced.” He noted that misuse rarely involves explicit, comic-book-style declarations of intent to build weapons for mass casualty. Instead, the company emphasized that the same data useful for developing pathogens could also be applied to creating vaccines or cures, making detection complex.
Anthropic highlighted five specific case studies involving biological misuse and six instances where Claude was utilized to develop software for conventional weapons, including firearms, missiles, armed drones, bombs, and their targeting systems.
The report also exposed the involvement of state-linked groups in cyber espionage and influence operations. A hacking group consistent with the Russia-affiliated Midnight Blizzard allegedly used AI to automatically detect when its malware was flagged by security defenses, subsequently rewriting code until it evaded detection. Additionally, an Iranian propaganda institution and China-based labs were identified as having misused the technology.
Anthropic clarified that none of the reported misuse cases involved its Claude Fable or powerful Mythos-class models, with the exception of one instance of distillation—a process used to train smaller models using larger, more expensive ones. The company stated it had shared intelligence with authorities and industry partners where appropriate.
These findings emerge amid growing anxiety within the tech industry regarding the pace of AI advancement. In early September, OpenAI chief scientist Jakub Pachocki published an article urging “voluntary slowdowns” in development, warning that the rapid rise of machine intelligence poses catastrophic risks and stating his belief there is a greater than 10% chance AI could “kill all humans” within the next decade.
Pachocki’s warning contributed to an open letter addressed to UK Prime Minister Andy Burnham, calling for a new multinational treaty on AI safety. In the United States, Democratic lawmaker Bernie Sanders has introduced legislation to ban AI superintelligence and impose a temporary pause on advanced AI development.
Does anyone else find it strange that the most dangerous misuse attempts targeted the smaller, faster models rather than the heavy hitters? Seems like a loophole worth investigating.
A Russian group using AI to rewrite its own malware? That is terrifying. If they can evade detection this easily, what else are they automating?
It is unsettling that the same data helps create cures and weapons. The nuance here makes prevention infinitely harder than just blocking bad actors.