Implementation guide
OWASP AISVS v1.01
Monitoring, Logging & Anomaly Detection · Detection and Alerting
Alert on jailbreak and injection attempts
Verify that the system detects and alerts on known jailbreak patterns, prompt injection attempts, and adversarial inputs.
OWASP AISVS v1.01, requirement 12.2.1
In plain English
Detect known jailbreak patterns, prompt injection attempts and adversarial inputs, and alert someone when they occur. Blocking an attempt is the first step. Knowing an attack is in progress lets you respond before one gets through.
What good looks like
- Detections from input and output screening generate security events.
- Alerts go to a person or team who is expected to act, with a severity that reflects volume and target.
- Detection rules are updated as new patterns appear.
- There is a written response for a sustained attack, such as suspending the account.
How teams usually do it
- Send guardrail and classifier events to your SIEM, and alert on thresholds: repeated attempts by one user, a spike across users, or any attempt against a high-privilege agent.
- Subscribe to sources of new attack patterns and add them to the detection set.
Suggested evidence
- Alert rules.
- Recent alerts and how each was handled.
- The response procedure.
Common mistakes
- Blocking attempts silently, so nobody sees an attack under way.
- One alert per blocked prompt, which buries the team until they mute the rule.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.