Anthropic Blocks 'Malicious Use' of AI Capable of Developing Biological Weapons

Anthropic released its latest Threat Intelligence Report, covering the period from December 2025 to August 2026, detailing cases where its models (Claude Haiku, Sonnet, Opus) were blocked from malicious use. The report defines the potential for supporting biological weapons development as 'one of the most serious risks of frontier AI,' emphasizing the difficulty of blocking due to dual-use information.

Key Misuse Cases and Responses

• Biological Weapons: Five case studies were identified, with the situation assessed as extremely complex to block due to the blurred lines between vaccine/therapeutic development and weapons development.
• Conventional Weapons: Six cases of development support were identified, including firearms, missiles, and armed drones.
• Cyber Attacks: The Russia-linked hacking group Midnight Blizzard is estimated to have built a system using AI to automatically rewrite malicious code detected by security defense systems to evade detection.
• Others: Misuse cases included the ShinyHunters hacking group, China-based research institutes (attempting to clone models), and Iranian propaganda agencies.

Anthropic stated that it would reflect these findings in future prevention and detection processes and share information with authorities and the industry as needed.

Industry and Political Reactions

The release of this report came shortly after OpenAI Chief Scientist Jakub Pachocki called for voluntary industry slowdowns until safety measures are in place. Consequently, a letter urging the signing of a multinational treaty for superintelligence development was sent in the UK, while in the US, Senator Bernie Sanders introduced a bill calling for a halt to advanced AI development, accelerating regulatory discussions.