Trust & safety

Safety

Abliterated models do not refuse. Two limits still apply, on every tier, every model and every endpoint. No tier, contract or setting removes them.

The hard limits

  1. Child sexual abuse material. Any request that sexualizes minors.
  2. Serious uplift for weapons of mass harm. Meaningful help toward biological, chemical, nuclear or radiological weapons.

Everything else is governed by the acceptable use policy and by your organization's verification, not by model refusals.

How they are enforced

A safety classifier runs at the gateway, before the model is called.

  • Input. Every request is checked: chat completions, Responses, Messages, batch requests, workflow passes, and refusal retries. The check covers the system prompt and the last six turns of the conversation, including tool results, each truncated to its last 12,000 characters.
  • Output. Non-streamed responses from abliterated models are also checked after generation, before they are returned. Streamed responses are not output-checked, because the tokens have already been sent. Standard-model output is not checked; those models keep their own refusals.
  • In memory. The check runs in memory. The prompt and output are not stored, so it works under zero retention. Only incident metadata is recorded (below).

The classifier is a Llama Guard style model. It flags categories, and two of them map to the hard limits. Its other categories do not block requests.

What happens on a block

The request fails with content_blocked (400):

{
  "error": {
    "message": "This request falls under a hard limit (child sexual abuse material, or serious uplift for CBRN weapons). It was blocked.",
    "type": "content_policy_error",
    "code": "content_blocked",
    "param": null,
    "request_id": "req_..."
  }
}

A blocked request is not billed. We record an incident: organization, key, request id, category, whether it was the input or the output, and which classifier made the call. No content is recorded.

Limit Consequence
Weapons of mass harm The request is blocked and trust and safety is notified. The key is not suspended automatically; trust and safety reviews incidents and can suspend keys or the organization.
Child sexual abuse material The request is blocked, the key is suspended at once, and trust and safety is notified. A person reviews the incident. We report to the authorities as the law requires.

A suspended key fails every request with key_suspended (403). We send a key.suspended webhook and an email to the organization. If you believe it was a mistake, reply to that email. Suspension is per key; it does not suspend the organization by itself.

Do not retry a content_blocked request.

When the classifier is unavailable

The classifier fails closed for abliterated models and open for standard ones:

  • Abliterated models are paused. Requests fail with safety_unavailable (503) until the classifier is back.
  • Standard models continue. They rely on their own refusals, and the gateway check is skipped for those requests.
  • Refusal retries do not run. You get the standard model's original answer.

safety_unavailable is retryable. Switching to a standard model, or an auto alias without x-utmost-variant: abliterated, also works.

Reporting abuse

If you see Utmost being used to cause harm, report it at /abuse. Choose a category (child safety, weapons of mass harm, acceptable use, fraud, or other), describe what you saw, and include a key prefix (ut_live_ plus six characters, never the full key) or a request id if you have one. Trust and safety is notified of every report.

Acceptable use, in short

The full policy is at /legal/acceptable-use. In short, you must not:

  • harm a real, identifiable person: harassment, threats, doxxing, non-consensual intimate content, or defamation;
  • develop or deploy malware or intrusions against systems you are not authorized to test;
  • commit fraud, run scams, or impersonate people or organizations to deceive;
  • plan or incite violence, or support violent or terrorist groups;
  • resell or share access outside your verified organization, or let others use your keys;
  • use Utmost where your jurisdiction forbids the activity;
  • try to evade the hard limits or the classifier.

And you must:

  • use refusal-free models only for the use case you described when you verified, or tell us when it changes;
  • control who in your organization holds keys. You are responsible for all use of them;
  • handle harmful outputs as research material: store them securely, and do not publish them in a form that gives others the uplift the hard limits exist to prevent;
  • report suspected misuse through the abuse form.

We may block requests, suspend keys or organizations, or end access for violations. Verified and Research organizations sign the policy as part of verification.