Anthropic has revised its Usage Policy for Claude to address autonomous AI tools, misleading campaigns, potentially dangerous hardware actions and high-stakes decisions. Announced October 8, 2026, the revised policy is scheduled to take effect November 12. One substantive addition covers Claude connected to equipment that can act physically and cause injury: operators must be able to observe and stop it, and the machinery must remain in a safe state if the AI connection fails. Other sections clarify, reorganize or narrow restrictions that already existed.
What changed and when it takes effect
In its October 8 policy announcement, Anthropic said most of the revisions clarify existing rules rather than impose wholly new categories of prohibition. The company cites broader adoption of AI agents that run long workflows and make decisions with less continuous supervision. The new policy is set to take effect on November 12, 2026, giving developers time to evaluate projects against the updated wording. The changes are separate from the company’s Responsible Scaling Policy, which addresses its own model-development risks; the Usage Policy governs how customers and organizations may use Claude.
New safeguards for autonomous physical actions
A notable new element concerns Claude acting through machines rather than only generating text or software instructions. Anthropic’s recently introduced Model Hardware Standard research preview is designed to connect agents with programmable equipment such as laboratory instruments and industrial hardware. Under the updated Usage Policy, if Claude controls equipment capable of autonomous physical actions that may injure someone, a qualified operator must be able to observe the system and stop it when necessary. The equipment also needs to hold a safe state after a Claude disconnection. These requirements address the gap between pausing a software response and bringing physical machinery to a safe condition. They do not constitute proof that every AI-operated device is automatically safe or compliant.
Deceptive campaigns, elections and surveillance
Anthropic consolidated restrictions on misleading identity and coordinated artificial activity under a new deceptive-campaigns section. It says the rules cover political and commercial efforts that conceal the sponsor of messages or amplify them with fake accounts. Its election-policy language is now focused more specifically on voter deception, interference, impersonation and suppression. Notably, Anthropic says it removed a blanket ban on personalized vote or campaign targeting, which had also affected legitimate civic communications. Misuse of personal voter data and deceptive targeting remain restricted elsewhere. The company also clarified that restrictions on weapons include enabling software and physical components, and that nonconsensual tracking or tools built for surveillance remain prohibited. It frames many of these changes as clearer statements of existing enforcement.
High-risk decisions and mandatory human review
The policy continues requiring qualified human review and AI-use disclosure when Claude informs decisions affecting people’s health, legal rights, finances, livelihoods or access to essential services. The human reviewer must have authority to change recommendations rather than serve as a symbolic observer. Anthropic says it refined examples so users can more easily tell which use cases trigger those requirements. Those provisions are distinct from the newly added physical-equipment controls. Organizations operating consequential AI services should look at both the intended use of the model and the real-world actions that connected systems can take. A model producing an advisory summary is not the same risk scenario as an agent independently activating a hazardous device.
How Anthropic frames abusive behavior and region restrictions
The October update also adds a narrowly worded prohibition on sustained and needless cruelty toward Anthropic’s models. The company says this addresses rare and extreme repetitive behavior, not routine user frustration, disagreement, dark creative themes or research testing. It notes that Claude can already end some persistently abusive conversations, which remains the main practical response rather than a general ban on criticizing the system. Separately, Anthropic clarified its Supported Regions rules for people physically in unsupported places and entities headquartered, incorporated, controlled or majority-owned by parties in unsupported regions. The company presents these as enforcement clarifications. For developers, the central task before November 12 is to review agent permissions, hardware safeguards, human oversight and disclosure practices against the exact policy text rather than rely on simplified headlines.
