Claude’s Guardrails End at Claude

Claude’s guardrails end at Claude. That sounds obvious, but Anthropic’s argument about distillation keeps trying to make its product boundary sound like a public safety boundary.

I use Claude Code. Most of the time, it is useful. Sometimes it refuses to finish ordinary work in a repository, and I cannot tell what threshold it believes I crossed. I have never tried to break into a system or bypass a security boundary.

That refusal protects Anthropic’s product policy. The underlying capability remains available elsewhere.

Open markets bypass closed models

Anthropic describes distillation as an attack: a competitor trains a less capable model on the outputs of a stronger one, acquiring expensive capabilities without bearing the original training cost. It has every right to detect that behavior and block it on Claude. Anthropic says exactly that.

The company’s commercial interest is direct. A frontier model costs billions to train. A competitor that can learn from its outputs reduces the lead Anthropic is selling through API access and subscriptions.

Anthropic also gives a safety argument. Its models restrict assistance with harmful activity. A laboratory that distills those capabilities into another model can remove the restrictions. That concern is real inside Claude’s own distribution channel.

The argument loses force once the capability exists elsewhere.

Anthropic can keep its weights closed while another company releases open weights. US labs can accept tighter limits while another country chooses a different policy. An open-weight model can matter without being free. A costly model remains out of reach for most people, yet companies and states will pay for a capability they find valuable.

No one can make Anthropic release Claude’s weights. Withholding them can delay the spread of capability. It cannot prevent it.

The cost leader changes the safety story

A Chinese lab can change this market with a model that is good enough, cheap enough, and available under looser constraints. Winning every benchmark is optional.

That is already a powerful route to adoption. Cost-sensitive developers choose the cheaper model. Companies with different risk tolerances choose the model that lets them run it privately. Researchers and governments pick the system they can inspect, modify, and keep running without an American company deciding what their work permits.

Distillation shortens that route. Cheaper chips, better training data, and a willingness to absorb legal or political risk shorten it too. Anthropic’s API controls govern one route. They do not create a global control point.

I want Anthropic to be honest about what its restrictions accomplish. They make Claude a safer service for customers who use Claude. Harmful actors can choose another sufficiently capable model.

Constitutional AI is a product feature

Constitutional AI gives Anthropic a way to train and operate Claude around a declared set of principles. That can make Claude more predictable. It can make Anthropic more comfortable selling Claude to companies like mine.

It also leaves the market unchanged outside Anthropic’s servers. Another model can use a different constitution, a weaker one, or none at all. Anyone with open weights can fine-tune away the provider’s preferred behavior.

Constitutional AI has real product value. It governs Anthropic’s service, not the wider model market.

This distinction matters when Anthropic pushes for policy around model access, weights, or distillation. Safety rules have to work against the actor who chooses another model, another jurisdiction, or a private deployment. Rules that only bind companies already willing to follow them become a cost of doing business for compliant labs and an advantage for everyone else.

Regulate the harm

I would focus policy on concrete harmful acts and deployments: malware operations, fraud, illegal surveillance, and systems built to select or attack people without meaningful human control. The law already knows how to regulate conduct. It can improve enforcement against people and organizations that use models to cause measurable harm.

Model weights create a harder problem because they cannot be recalled after release. A broad campaign against open weights punishes labs that publish openly and rewards actors who ignore the rule or operate somewhere else.

A delay can matter when a model crosses a genuine dangerous-capability threshold. Anthropic should make that case with evidence, define the threshold, and accept independent testing. Its own policy has argued for independent evaluation rather than company-defined rules for open models. That is the right instinct.

I will keep using Claude Code because its model and harness are useful. A refusal banner governs one product surface. Public safety needs controls that still work after the customer opens another tab.