Anthropic updates Claude rules for agents and cyber risk

AI is no longer just typing back answers in a chat box. Over the past year, it got hands, learned how to run terminal commands, and started clicking around our desktops.

Anthropic officially published an updated Usage Policy to address these exact shifts. The new rules take effect on September 15, 2025, bringing clear guardrails for autonomous agents while loosening overly strict bans on political discourse.

Drawing lines for agents and political research

The biggest changes zero in on cybersecurity and agentic tools. With products like Claude Code and Computer Use now live, Anthropic models power a big chunk of modern coding agents.

Following its March 2025 threat intelligence report, the company added dedicated rules barring unauthorized infrastructure compromises, malware generation, and automated attacks. However, authorized vulnerability research conducted with system owner consent remains explicitly permitted.

Anthropic also walked back its old blanket ban on political content. Previously, almost any campaign or lobbying material was blocked, which ended up frustrating users trying to do civic education or public policy analysis.

Under the new guidelines, legitimate political discourse and research are allowed. The ban now targets deceptive campaigns, efforts to disrupt democratic elections, and intrusive voter targeting.

There is also a notable clarification for high-risk domains like law, finance, and employment. While human oversight and AI disclosure requirements still apply when dealing with everyday consumers, Anthropic made clear that purely B2B operations do not face these consumer-facing compliance burdens. Meanwhile, the ban on law enforcement surveillance, tracking, and biometric profiling remains firmly in place.

Why it matters

We think this policy revamp shows Anthropic maturing past the paranoid stage of AI safety into practical governance.

For enterprise teams and researchers, it removes massive headaches. Policy analysts can finally run civic texts without triggering random safety blocks, and B2B developers building workflow tools get relief from consumer-grade human-in-the-loop demands.

Setting rules for text generation is easy, but policing autonomous agents like Claude Code is an entirely different beast. Drawing a clean technical line between automated penetration testing and unauthorized infrastructure compromise is notoriously difficult in practice.

Anthropic is making a smart bet by treating enterprise tools differently from consumer bots. But as agentic capabilities grow, keeping rogue actions contained will require far more than just updating terms of service.