Implementation guide
OWASP AISVS v1.01
Model Lifecycle Management & Change Control · Model Validation & Testing
Test the model before each deployment
Verify that models undergo automated input validation testing, safety evaluation testing, and output sanitization testing before deployment.
OWASP AISVS v1.01, requirement 3.2.1
In plain English
Before any model goes live, run automated tests on three things: how it handles malformed and hostile input, whether it stays within your safety rules, and whether its output is cleaned before other systems use it. If a test fails, the deployment should stop.
What good looks like
- A standard test suite runs in the pipeline for every model, including new versions and fine-tunes.
- The suite covers hostile input (prompt injection, jailbreaks, oversized and malformed requests), safety behaviour and output handling.
- Pass thresholds are written down, and a fail blocks deployment.
- The suite grows after every incident and red-team finding.
How teams usually do it
- An evaluation stage in the deployment pipeline using an open-source scanner or framework, for example garak, PyRIT, promptfoo or Inspect.
- A team-owned set of regression prompts drawn from past incidents and public jailbreak collections.
- Tests run against the full application (prompt, filters and model together) as well as the bare model.
Suggested evidence
- The test suite and its pass criteria.
- Results for the model currently deployed.
- A pipeline run where a failing result stopped a release.
Common mistakes
- Testing task accuracy only, with no safety or abuse tests.
- Running the tests by hand when someone remembers.
- Testing the bare model and skipping the system around it, where output-handling bugs live.
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.