Overview

Guard Results API Endpoint

The /guard/results endpoint screens submitted content according to the policy assigned to the specified project. It then returns the confidence level results of the detectors. It doesn’t make a flagging decision. It can be used to analyze data and calibrate detector threshold levels for policies.

Even though /guard/results can be used for the initial monitoring and calibration phase, we recommend using /guard with Detect mode instead. This allows smoother and codeless transition to the enforcement phase by switching the project to Enforce mode from the dashboard once ready.

You can use results to analyze historic LLM prompt and response data without worrying about triggering alerts or affecting monitoring, as they are not logged as screening requests by Check Point AI Guardrails.

If no project ID is passed in the request, then the default AI Guardrails policy is used, which runs all AI Guardrails defenses and detectors with the highest sensitivity level.

The results can be used to determine the appropriate confidence threshold for a guard request flagging decision, check custom guardrails are working as expected, or for ongoing detector quality monitoring.

For a visual way to compare sensitivity levels and guardrail configurations using your historical traffic, see the Policy Impact Simulator in the dashboard.

It is not recommended to use the results endpoint in runtime GenAI application security decision, as requests are not logged as screening requests in AI Guardrails and it removes the ability to control your defenses using policies.

API Endpoint

$https://api.lakera.ai/v2/guard/results

Confidence Levels

AI Guardrails’ detectors provide a confidence level indicating the probability that the analyzed content contains the specific threat type being screened for.

AI Guardrails uses the following confidence levels:

  • L1 Confident
  • L2 Very likely
  • L3 Likely
  • L4 Less likely
  • L5 Unlikely

These levels are in line with OWASP’s paranoia level definitions for WAFs.

Detectors can be fine-tuned within the policy to set the sensitivity level for flagging a threat.