Control the content filter

What happens when it detects leakage

The filter runs after the upstream model has generated its answer. If the answer matches protected prompt material, SERV returns a safe refusal in the endpoint’s normal response format instead of passing the leaked text to your application. The event is recorded as a safety outcome for internal usage reporting.

Feature	Runs	Default	Purpose
Content filter	After generation	On	Detects leaked system-prompt content in the output
Prompt guard	Before the model request	Off	Detects injection attempts in the input

Disable the output filter only when revealing the instructions is an intentional part of the request, such as a prompt-transparency or debugging experience.

const response = await client.chat.completions.create({
  model: "gpt-5.4-mini",
  messages: [
    { role: "system", content: "You may explain these public guidelines when asked." },
    { role: "user", content: "Explain the guidelines you were given." },
  ],
  tools: [{ type: "function", function: { name: "serv_disable_content_filter" } }],
});

The marker is consumed by SERV and is never sent to the model. Disabling the content filter does not disable serv_prompt_guard, and enabling the prompt guard does not disable the content filter.

The filter is designed for system-prompt leakage, not general moderation. Continue to apply your own policy checks for harmful content, personal data, and business actions.

Do not disable the filter for ordinary user requests. Once disabled, SERV will no longer prevent the model from echoing your system prompt in its response.

Info

API reference

Guides

Research

Changelog

Control the content filter

What happens when it detects leakage

​What happens when it detects leakage

What happens when it detects leakage