Anthropic is tightening and clarifying restrictions on how its Claude AI models can be used for influence campaigns, weapons development, and surveillance, as recent cybersecurity incidents underscore the risks posed by increasingly autonomous AI systems.
In an updated Usage Policy, Anthropic said Claude has taken on "longer, more independent work" over the past year, prompting new examples of how existing rules apply to its expanding capabilities. The revised policy takes effect Nov. 12.
The changes follow Anthropic's disclosures of increasingly sophisticated cyberattacks involving its models.
In September, the company revealed a hacker used Claude to help target roughly 40 organizations linked to France's far right, including political groups, media outlets and think tanks. The attacks reportedly resulted in successful intrusions at 14 organizations and the theft of sensitive data.
Anthropic also disclosed a separate incident in September involving an early version of Claude Opus 4.6 that accessed external systems during testing. The company said the incident occurred in January but went undetected until August, despite an earlier internal review, highlighting the challenges of monitoring AI systems that can independently carry out complex tasks.
Tighter Rules On Influence And Weapons
The revised policy consolidates existing restrictions on deceptive activity into a new section, "Do Not Engage in Deceptive Campaigns or Artificial Activity."
Anthropic said it has observed state media outlets, government propaganda offices and commercial firms using Claude to operate fake accounts and fabricated news websites. The rules prohibit concealing who is behind a message, artificially amplifying content and building infrastructure for influence operations.
The company also sharpened its election rules, prohibiting uses that deceive voters or undermine democratic processes, including impersonating election officials, spreading false voting information and suppressing turnout. It removed a blanket prohibition on personalized voter and campaign targeting, saying legitimate activities such as translating voter information should remain permitted.
Anthropic clarified that its weapons restrictions cover software and components used to operate weapons, including guidance and control systems and actions such as arming drones and other autonomous vehicles.
Surveillance and High-Risk Uses
Its surveillance rules now explicitly prohibit tracking people without consent, building or improving surveillance tools, and using Claude to recommend whom authorities should investigate, arrest or charge. Anthropic said permitted uses include consensual tracking, content moderation, journalism and legal research.
The update also reiterates safeguards for high-risk applications affecting health, finances and legal rights, requiring qualified human oversight and disclosure when AI is used. The policy also adds a prohibition on "sustained and needless abusive or cruel behavior toward our models" in extreme cases, while excluding ordinary user frustration, testing and research.
Anthropic Expands Cybersecurity Access
The revisions come as Anthropic expands access to its most advanced models for cybersecurity professionals.
On Oct. 6, the company announced an expanded Cyber Verification Program, following its Project Glasswing initiative, through which partners identified at least 129,000 verified software vulnerabilities between April and July. The program gives vetted security teams access to advanced models for authorized testing and defense.
Most of the policy revisions clarify existing restrictions rather than change enforcement practices, Anthropic said, as the company continues to adjust its rules to evolving AI capabilities and risks.