AI biological weapons risk escalated after scientists attempted to deploy frontier artificial intelligence models for dual-use pathogen research, United States AI safety firm Anthropic confirmed on Thursday, according to Reuters. The San Francisco-based developer documented five distinct case studies between December 2025 and August 2026 where users sought technical assistance that could advance biological weapons development.
The company disrupted the operations before physical biological agents were generated.
According to the threat intelligence findings, one user attempted to draft a research grant application focused on gain-of-function modifications to the chikungunya virus. The proposed engineering aimed to elevate viral transmissibility and bypass immune responses in host organisms. Another detected case involved avian influenza research structured around mammalian adaptation experiments, raising biosecurity alarms across automated classifiers.
Jacob Klein, head of threat intelligence at Anthropic, stated that detecting bioweapons research is an extraordinarily nuanced operational challenge. Valid civilian research into vaccines, therapies, and infectious diseases inherently overlaps with the exact biochemical sequences required to synthesize dangerous biological weapons. Anthropic noted that the individuals identified were active researchers, and investigators could not definitively determine whether malicious intent was present.
What remains clear is that computational safeguards face continuous circumvention attempts.
The investigation revealed that threat actors utilized intermediary relay platforms and evasion tactics after encountering automated refusals. Beyond biological risks, Anthropic disrupted six operations where actors in Russia, China, and Yemen deployed the Claude model to generate targeting and control software for conventional armaments, including missiles, armed drones, and munitions. The company also reported state-backed cyber espionage workflows and propaganda efforts targeting institutional networks.
The publication comes amid broader warnings across the technology sector, with Google disclosing earlier this week that an individual attempted to generate synthesis protocols for weaponized biological agents using Gemini. United States lawmakers have renewed calls for strict regulatory caps, with Senator Bernie Sanders advocating a pause on advanced frontier models. Anthropic confirmed all implicated accounts were terminated and stricter classifiers were deployed across newer model generations.
