Anthropic Blocks Claude Use Tied To Biological-Weapons Risk
BBC reports that Anthropic disrupted attempts to use Claude for biological-weapons support, alongside cases involving conventional weapons, cyber operations and surveillance.
Anthropic blocked several attempts to use Claude in work that could support biological weapons development, the BBC reported, turning the company's latest threat intelligence report into a test of how AI labs police high-risk model use.
The cases sit inside a wider misuse review covering activity between December 2025 and August 2026.
The company grouped the activity across government-linked attackers, criminal operators, commercial spyware firms, propaganda bodies and politically motivated users that tried to turn its systems toward harmful work.
The biological-risk section carried the highest stakes.
Five case studies involved actors using Anthropic models in ways that could support biological weapons development.
The company called biological misuse one of the most serious risks for frontier AI models and warned that, without safeguards, such capabilities could have catastrophic consequences.
The same technical material can also support legitimate medicine.
Information useful for a biological weapon can overlap with material needed to develop a vaccine or disease treatment, Anthropic warned.
That dual-use problem makes the enforcement question less simple than a user openly asking to build a weapon.
Jacob Klein, Anthropic's head of threat intelligence, told the New York Times that the situation was “incredibly nuanced”.
The source quoted him saying investigators were not seeing a comic-book-style request from someone saying they wanted to build a biological weapon to kill everyone.
The report also covered conventional weapons and cyber operations.
Six cases involved Claude being used to write software connected with guns, missiles, explosive devices, armed unmanned aircraft and the targeting or control systems behind those weapons.
Other cases included fake dating apps, hotel Wi-Fi scams and surveillance tools intended to identify dissidents.
Cybercriminal and state-linked activity formed another strand of the review.
The BBC account placed Claude use inside activity tied to a Russian cyber-espionage operation and an Iranian propaganda institution.
One hacking group, whose work matched patterns associated with Midnight Blizzard, allegedly used AI to automate a cycle in which hostile code was checked against security tools, altered after detection and tested again.
Anthropic incorporated the findings into internal processes to better prevent, detect and disrupt similar activity.
The company also shared intelligence with authorities and industry partners where appropriate.
The warning lands during a broader policy dispute over advanced AI.
Google disclosed earlier in the week that a person had tried to use Gemini to obtain a complete technical guide for synthesising weaponised biological agents.
US Senator Bernie Sanders has proposed temporarily halting advanced AI work while barring superintelligence development, while President Donald Trump framed the race differently, warning that the United States would be in a bad position if it did not win AI.
A second industry warning gave the debate more force.
OpenAI chief scientist Jakub Pachocki called for voluntary slowdowns until safeguards are set, while an open letter to UK Prime Minister Andy Burnham sought a multinational treaty on safe AI development.
For AI developers, the operational result is no longer just model capability.
The biological, weapons and cyber examples make detection, escalation, user controls and intelligence sharing part of the product's security boundary.
