Implementation guide OWASP AISVS v1.01

Model Behavior, Output Control & Safety Assurance · Output Format Enforcement

Don't act when the AI refuses or fails

AISVS 7.1.3 Level 2

Verify that, when an application uses an AI response to decide whether to perform an action, it does not perform that action if the AI refuses the request, reports a failure, or returns a response the application cannot correctly read or understand.

OWASP AISVS v1.01, requirement 7.1.3

In plain English

If your software uses the model's answer to decide whether to do something, it must do nothing when the model refuses, reports an error or returns something the code cannot read. The safe default is no action. Systems that fail open here approve things because a response was garbled.

What good looks like

  • An action runs only on a clear, valid, positive result.
  • A refusal, an error, a timeout, a truncated response or an unparseable one leads to no action.
  • The code checks for approval explicitly. It does not check for the absence of a refusal.
  • These cases are logged and counted.

How teams usually do it

  • An explicit decision field in the output schema, with a fixed set of values such as "approve", "deny" and "unsure". Only "approve" proceeds.
  • Code written so the default branch does nothing, and every exception path ends in "no action".
  • Reading the provider's refusal and finish-reason fields where they exist, in addition to the content.

Suggested evidence

  • The decision logic in code.
  • Tests with refusals, errors, empty responses and garbled output, each showing no action taken.
  • Log counts of declined or unreadable decisions.

Common mistakes

  • Treating any non-empty response as approval.
  • Searching the text for the word "yes" or "approved", which also matches "I cannot say this is approved".
  • Retrying on failure until the model eventually says yes.

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.