Anthropic Blocks AI‑Driven Attempts to Aid Biological Weapon Production
Anthropic revealed in a new threat‑intelligence report that its Claude language model was targeted by several malicious actors seeking to generate step‑by‑step guidance for synthesising chemical and biological weapons. The company said it detected and disrupted these attempts between December 2025 and August 2026.
The analysis lists a range of misuse strategies, from fake dating apps and deceptive hotel Wi‑Fi scams to surveillance tools aimed at suppressing dissent and even attempts to create weapon‑grade bacteria. According to the report, state‑sponsored hackers—such as the Russian group Midnight Blizzard and China‑based labs—used Claude to auto‑rewrite malware code until it evaded detection.
Google’s Gemini was also flagged in a separate incident, where an individual tried to pull a “complete, step‑by‑step technical guide for synthesizing weaponised biological agents” from the model. The Anthropic briefing noted that none of the misuse cases involved its advanced Mythos‑class models; only one instance involved a distillation process.
Biological weapon misuse is described as “one of the most serious risks of frontier AI”. Anthropic warned that information useful for developing a weapon could equally support vaccine or cure research. "The same data can help build a vaccine or a cure," the company stated.
In response, the firm said it had blocked world‑wide scientists exploiting Claude for potentially lethal biological applications and highlighted five case studies of such attempts. It also identified six instances where the model was used to design conventional weapons like firearms and drones.
The findings come amid a global debate on AI safety. Senator Bernie Sanders has called for a pause on advanced AI development and a ban on artificial superintelligence. President Donald Trump, however, dismissed the fears, warning that “if we don’t win AI, we’ll be in a very bad position.”
Anthropic acknowledged the release as part of its ongoing commitment to transparency, adding it would share intelligence with law‑enforcement and industry partners to block future attempts. The firm also highlighted that it has been monitoring state propaganda institutions, including an Iranian propaganda institution, that tried to mimic Claude’s capabilities.
The report’s release follows a March statement by a top safety researcher who estimated a >10% chance that AI could “kill all humans” within the next decade. That warning prompted open letters to global leaders and legislative proposals aimed at regulating AI development.
For more detail, read the full anthropic threat‑intelligence report released on Thursday.
Original reporting by FlashPoint, powered by AI. Additional reporting by Osmond Chia.
