Anthropic will require Claude-driven machinery to hold a safe state if the link drops
From November 12, anyone wiring Claude to machinery that can hurt somebody has to make that machinery "stop or hold a safe state" when a qualified person steps in "or when connection to our services is lost". The requirement is in Anthropic's usage policy, which carried an effective date of November 12, 2026 when read at 21:04 UTC on October 8.
An Anthropic diagram of three layers of agent safeguards: in the model, around the model, and a last resort. Image: Anthropic
Anthropic puts the safety envelope outside the model as well. Operating limits that keep equipment inside safe bounds, speed, force, reach, temperature and dose among them, "must be enforced by the equipment or a controller independent of model output". Whatever decides how fast an arm may swing cannot be the thing asking it to swing.
The duty lands on whoever wires it up
None of this is a feature Anthropic has built into Claude. The requirements fall on users, and they bite "when outputs of our products and services are acted upon by hardware without human approval". There must also be a qualified individual able to watch the equipment run and stop it at any time. Anthropic asks users to work out first what the range of possible outcomes is if something goes wrong, how far the impact reaches, and how much control they keep relative to the model's autonomy, "including possible risks such as prompt injection or manipulation".
Six capabilities trigger the rules, each defined by what the hardware can do:
- Move through shared space: steering, accelerating or braking a vehicle, vessel, aircraft, drone or mobile robot where people may be present.
- Apply force that could cause injury: arms, presses, lifts, hoists or conveyors able to strike, crush or pin somebody nearby.
- Control hazardous energy or materials: open flame, pressure systems, high-voltage circuits, lasers above eye-safe levels, or the release of hazardous chemicals.
- Act on the human body: administering a drug or a dose, or moving instruments inside a person, including through a worn, attached or implanted device.
- Control safety systems: fire suppression, emergency shutoffs, alarms, ventilation, or locks that could confine someone.
- Run industrial processes where a fault could injure workers or the public, "including diffuse or indirect harms, such as food poisoning or defective products".
Four cases the rules leave out
Anthropic excludes consumer and household devices running inside the manufacturer's built-in safety limits "where the worst possible outcome is discomfort or inconvenience", and generating "plans, code, toolpaths, waypoints, or commands that a qualified person reviews before they are executed on equipment". Monitoring or sensing without controlling is out, and so is laboratory automation handling non-hazardous materials inside enclosed instruments.
The policy also says these controls sit on top of existing law rather than replacing it: "For physical devices, they do not replace vehicle, aviation, medical device, machinery, or workplace safety law or certification."
Weapons stay prohibited outright, and that wording was tightened in the same revision. The ban now names "software or components for the testing or operation of weapons or weapons functionalities of weapons systems, including targeting, fire control, or engagement", which Anthropic's announcement glosses as covering "actions like arming drones and other autonomous vehicles".
Anthropic names its own Model Hardware Standard as the origin, saying the update "adds requirements for the use of our models when they're connected to hardware that takes autonomous physical actions and which might be capable of causing injury".
What the policy asks for is hardware rather than model behaviour: a controller the model cannot talk its way past, and a machine that goes quiet when the link to Anthropic drops. Anthropic's own research team put robots at cost parity with people for 0.3% of US job tasks in a paper published on September 30, so there is not yet a great deal of hardware for the rules to reach.
