How to meet it
OWASP AISVS v1.01
Adversarial Robustness · Membership-Inference and Model-Inversion Mitigation
Don't reveal attributes the model inferred
Verify that model-inferred sensitive attributes are not directly returned in outputs.
OWASP AISVS v1.01, requirement 11.2.1
In plain English
If the model can guess sensitive things about a person, such as health, religion or sexuality, do not show those guesses in its output.
What good looks like
- Sensitive attribute categories are listed.
- Outputs are checked for inferred values in those categories.
- Such inferences are removed or blocked.
How teams usually do it
- Output filters for sensitive categories.
- Prompt and product design that avoids asking the model to infer them.
Evidence an auditor would accept
- The category list and filter tests.
Common mistakes
- A profile feature that "helpfully" guesses personal characteristics.
Not sure which requirements apply to your system? Take the assessment.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.