Anthropic mandates safety limits for Claude-controlled hardware

New mandates for physical AI

Starting Thursday, Nov. 12, Anthropic is implementing a new High-risk Physical Actions section in its Usage Policy. This update specifically targets anyone using Claude to drive equipment that could injure a person without constant human approval. The policy requires that a qualified individual must be able to observe the equipment and stop it at any time.

The new rules are designed to address the increasing independence of Claude in physical tasks. Anthropic linked these updates to its Model Hardware Standard research preview, which explores how AI interacts with physical machines. The policy applies to users of Claude directly as well as customers using the model through Amazon Web Services' Bedrock.

While the policy sets these new requirements, it does not replace existing safety laws. Operators must still comply with established vehicle, aviation, medical-device, and machinery safety rules. The update serves as an additional layer of safety for AI-integrated hardware rather than a replacement for government certifications.

Independent safety controllers

One of the most significant requirements is that safety limits must be enforced by the equipment itself or an independent controller. This means that constraints on speed, force, temperature, or dose cannot rely on the model's output. This approach is a direct response to security research showing that LLM-controlled robots can be 'jailbroken' through prompt injections.

Because a model can be tricked into ignoring its own internal instructions, a separate controller provides a hard barrier. If the AI sends a command that exceeds safe parameters, the independent system must block it. This ensures that the machine's physical behavior remains within safe bounds regardless of what the AI requests.

Anthropic's Model Hardware Standard provides a framework for these drivers, but it is unclear if a limits file within that specific software stack qualifies as an independent controller. Operators should treat this specific technical requirement as an open question until Anthropic provides further clarification. For now, the burden of proof for independence lies with the equipment operator.

Handling connection losses

The policy introduces a specific requirement for what happens when the connection to Anthropic's services is lost. Equipment must be designed to stop or hold a safe state immediately if the link drops. This prevents a robot or machine from continuing its last command indefinitely without guidance from the AI model.

This requirement turns network stability into a critical safety dependency for AI hardware. Integrators must decide what the specific 'safe state' should be for their particular application. Depending on the hazard, a safe state might mean the machine stops entirely, holds its current position, or retracts a moving part.

Because the policy does not define how long a stalled link counts as 'lost,' operators must establish and test their own thresholds. This means that testing the hardware by intentionally cutting the connection is a necessary step for compliance. It ensures the machine behaves predictably during a cloud outage or a local network failure.

Scope and exclusions

The policy identifies six specific situations that trigger these high-risk rules. These include equipment moving through shared spaces, machines applying force capable of causing injury, and systems controlling hazardous energy or materials. It also covers devices acting on the human body, safety systems, and industrial processes where a fault could lead to injury.

Conversely, four types of use are excluded from these specific hardware mandates. Consumer and household devices are exempt if they operate within built-in safety limits and the worst outcome is only discomfort. Plans, code, or commands that a qualified person reviews before execution are also excluded from the hardware section.

Additionally, monitoring or sensing tasks that do not control any physical actions are not covered. Laboratory automation is also exempt as long as it handles non-hazardous materials inside enclosed instruments. These exclusions help distinguish between high-risk industrial AI and everyday consumer technology or human-in-the-loop workflows.

Broader policy changes

Beyond hardware, Anthropic updated its policy to include eleven high-risk recommendation areas. The policy also clarifies bans on weapons development and the integration of weapons onto drones or other unmanned platforms. It covers decisions regarding investigations, arrests, and sentencing to prevent the use of Claude in sensitive law enforcement actions.

The elections section was also narrowed to allow for legitimate civic work. Anthropic removed a blanket ban on personalized vote and campaign targeting to accommodate nonprofits and election officials. However, the policy still prohibits deceptive campaigns and the seeding of misleading content in AI-generated search or answer sources.

Finally, the policy addresses unauthorized access by banning the reselling or proxying of Claude subscriptions. This targets workarounds where users might route requests through overseas-registered accounts to bypass regional restrictions. These changes aim to ensure that Claude is used by authorized entities in compliance with Anthropic's regional availability rules.

Next steps for operators

Operators with Claude connected to any physical actuators should perform a series of checks before the November 12 deadline. They must map every Claude-to-hardware link against the six triggers and four exclusions. They also need to identify a qualified individual for each system and document why that person is capable of holding the stop role.

Technical testing is also required to ensure compliance with the connection-loss and independent controller rules. Operators should physically cut the connection during a test to confirm the equipment enters a safe state. They must also verify that safety limits hold even when the model sends a command that is out of the allowed range.

For those using Amazon Bedrock, continued use on or after the deadline will be considered an acceptance of these new terms. Anthropic has stated it will continue to seek input from policymakers and experts as the technology evolves. For now, the engineering and compliance burden for hardware safety rests primarily with the operators.

Compliance Checklist

Before the new rules take effect, organizations should conduct a thorough audit of their AI-integrated hardware. This includes identifying every instance where Claude interacts with a physical actuator. Each instance must be evaluated to see if it falls under the high-risk categories, such as moving through shared spaces or handling hazardous materials.

A key part of this audit is defining the human oversight role. The policy requires a qualified individual who can observe the equipment and intervene. Organizations need to formally designate these individuals and ensure they have the means to stop the machinery instantly. This human-in-the-loop requirement is a core pillar of the new safety framework.

Finally, the technical architecture must be verified to ensure safety limits are independent of the model's output. If a model's response is the only thing governing the speed or force of a machine, the system is non-compliant. Moving these limits to a separate controller or the hardware itself is the primary way to meet Anthropic's new standards.

Watch for clarification

One of the most important things to watch is how Anthropic defines the 'independent controller' for the Model Hardware Standard. While the standard exists, the policy doesn't explicitly say if a limits file loaded by the same software stack meets the requirement. Operators should wait for official guidance before assuming their current software architecture is sufficient for compliance.

Another area to watch is the definition of 'connection lost.' Since the policy leaves this undefined, different industries may adopt different timeouts for what constitutes a lost link. These choices will impact the reliability of AI-driven machines, as a robot that halts on every minor network blip might be less useful than one that waits for a few seconds.

Lastly, keep an eye on how Anthropic enforces the new 'cruelty' ban. While the policy mentions the ability of Claude to end conversations as a primary enforcement mechanism, it is unclear if repeated violations will lead to account suspensions or other penalties. The company has not yet clarified the specific consequences for users who repeatedly violate these behavioral guidelines.

Try it today

If you use Claude to control any hardware, perform a 'pull-the-plug' test: disconnect your internet or service link and verify that the machine immediately enters a safe state without further instructions from the AI.