Claude-connected equipment capable of injuring someone will need more than a safety instruction in a prompt. Anthropic’s revised Usage Policy requires a qualified operator who can observe and stop the equipment, a safe state if Claude disconnects, and operating limits enforced independently of the model’s output.
These hardware safeguards are among the most consequential changes in Anthropic’s policy announcement, published October 8, 2026. The revised rules take effect November 12. The update also prohibits extreme, repeated abuse directed at its models, consolidates restrictions on deceptive campaigns, narrows its election rules, and makes weapons-software and surveillance prohibitions more explicit.
The model-abuse clause is unusual, but it shouldn't obscure the operational changes. For API developers, agent builders, and organizations using Claude in consequential decisions, the update addresses how systems behave, who supervises them, and what infrastructure they are allowed to help build.
Hardware Safety Cannot Depend on Claude’s Output
The hardware provisions address models connected to equipment that takes autonomous physical actions and might cause injury. According to Anthropic’s Usage Policy, these deployments need three separate protections:
- A qualified operator must be able to observe the equipment and stop it when necessary.
- The equipment must be able to maintain a safe state if Claude disconnects.
- Operating limits must be enforced independently of model output.
Each requirement addresses a different failure mode. An operator provides supervision and intervention; a safe state addresses loss of the model connection. Independent operating limits constrain what the system can do even when Claude produces an unsafe instruction.
A system prompt telling Claude to respect a machine’s limits cannot, by itself, satisfy that last requirement. The safety boundary must remain effective regardless of what the model generates. Builders also need to consider what the equipment does when Claude becomes unavailable. Continued model access cannot serve as a safety assumption.
These deployment requirements make the equipment’s control design part of compliance. They provide no assurance that the model will always act safely: supervision, interruption, and safe operation cannot all depend on the same model whose actions they are supposed to constrain.
An agent builder’s review should extend beyond prompts and tool permissions to the entire path between model output and physical action, including what can prevent or stop a dangerous command.
