Anthropic Updates Claude Usage Policy for Agents and Politics

Anthropic has announced a comprehensive update to its product usage guidelines, set to take effect on September 15, 2025. The revised policy addresses the rapid rise of agentic AI technologies, such as the company's own Claude Code and Computer Use tools. Following findings from its March 2025 threat intelligence report on malicious Claude usage, Anthropic is explicitly banning the use of its models for malicious system, network, and infrastructure compromise. However, the company will continue to permit authorized cybersecurity activities, such as consensual vulnerability discovery.

The update also marks a shift in how Anthropic handles political content. Moving away from a blanket ban on lobbying and campaign material, the company is tailoring its rules to allow civic education, policy research, and political writing. It will still prohibit deceptive practices, voter targeting, and efforts to disrupt democratic processes. Additionally, Anthropic clarified its rules for law enforcement, maintaining bans on biometric surveillance and tracking while simplifying the language regarding permitted back-office and analytical applications.

For enterprise developers, Anthropic clarified that its high-risk safeguards—which require human oversight for legal, financial, or employment applications—only apply to consumer-facing deployments, not business-to-business interactions. The company is also adjusting the biological safety filters for its Claude Fable 5 model. This adjustment aims to minimize false positives, preventing the system from unnecessarily downgrading to a weaker model when processing safe scientific inquiries. Finally, Anthropic is establishing a $5 million grant initiative to support external studies on the relationship between AI and human wellbeing.