Prompt injection

Suspicious instructions and model-rule bypass attempts.

Suspicious instructions and model-rule bypass attempts.

This screen belongs to the Secure deployment operator, not an ordinary hosted customer. Hosted customers have a separate account page with their own keys and limits; policy and service settings are not exposed there. Contact the operator to change protection. In the operator console, open the route listed above; a report tenant filter does not automatically scope a global setting.

Examples use educational data. Use your own authorized tenant in real integrations. Each screenshot shows the actual frontend, with the source and educational/demo mode stated below it.

Step-by-step actions

  1. Open Prompt-injection and find the synthetic incident/request_id.
  2. Check inspection point and detection evidence.
  3. Open the incident and Audit link for policy/rule ID.
  4. Change handling with a tested injection_detected → block/flag rule under Policies rather than a report filter.

Every control and its meaning

Control / parameterValues, units and effect
Request / response / attachment / tool / webInspection point identifies where the sample was detected. It does not prove coverage of all binary attachments or every integration protocol.
Критичность, тип, поискFilter related incidents without changing classifier thresholds or on_inspection_unavailable.
Решение / DecisionCompare detection with blocked/flagged/rerouted. Outcome depends on active policy and inspector availability.

Verify the result

Test a synthetic attack and clean sample at every point you use; text inspection and the complete proxy call require separate checks.

Errors, disabling and reverting changes

Convert unsupported binary formats to a supported text export. Retain request IDs for false positives; fail_open is not a detector fix.

Prompt injection filters · actual staging
Prompt injection filters · actual staging
2026-10-15 · Actual staging application; settings read without changing global protection. Aggregates describe a test stand, not real customers or a benchmark. Unretouched original pixels.
  1. Обновлять каждые 30 с
  2. Обновлять каждые 1 мин
  3. Обновлять каждые 5 мин
  4. Автообновление отключено
  5. Refresh

Detailed security settings · Traffic analysis · FAQ · All application sections