Skip to main content
Planck can inspect user messages for prompt-injection and jailbreak patterns before a standard chat-completions request reaches a model provider.
Prompt security is an account and deployment capability. It is not enabled by selecting a particular model, and the public API does not currently expose per-request thresholds.

​
Request behavior

When prompt security is active for your traffic, Planck:
  1. Extracts text from user messages.
  2. Scores it with the configured prompt-attack detector.
  3. Records the detector result with the request’s security properties.
  4. Blocks the provider call with HTTP 403 when the configured block threshold is reached.
If the detector is unavailable or times out, the current gateway behavior is fail-open: the provider request continues and the failure is recorded operationally.

​
Blocked requests

A blocked request returns a standard gateway error response. Applications should handle the HTTP status and machine-readable error fields instead of matching the human-readable message.
The exact response fields can vary with the compatibility route. Treat 403 as a policy rejection and show a safe, application-specific message to the user.

​
Application guidance

  • Keep system instructions separate from untrusted user content.
  • Give tools the minimum permissions and validate every tool argument.
  • Require confirmation for destructive or externally visible actions.
  • Avoid putting secrets in prompts or model-visible tool output.
  • Log session and user identifiers so a rejected workflow can be investigated.
Prompt-attack detection is one layer in a broader security design. It does not replace authorization, input validation, output validation, or tool-level access controls. For account availability and security configuration, contact help@inquantum.ai.