Implementation guide
OWASP AISVS v1.01
Training Data Integrity & Traceability · Training Data Quality and Security Assurance
Remove disallowed content before training
Verify that disallowed content is detected and removed before training.
OWASP AISVS v1.01, requirement 1.3.4
In plain English
Find and remove content the model must not learn from before training starts. That covers illegal material, content your policy bans, personal data you have no basis to use, and leaked secrets such as passwords and API keys. Once it is in the model's weights, you cannot reliably take it out.
What good looks like
- A written list says which content categories are disallowed in training data.
- Automated filters run over all new data before it is used, including fine-tuning and feedback data.
- Removed items are counted and logged so you can check the filter is working.
- The filters are tested against known examples of each category.
How teams usually do it
- Secret scanners such as Gitleaks, TruffleHog or detect-secrets over code and text data.
- Content classifiers for the banned categories, and hash matching against lists of known illegal images (PhotoDNA is the usual example).
- Personal-data detection with masking or removal, as in the labelling stage.
Suggested evidence
- The disallowed-content policy.
- Filter configuration, and counts of items removed per run.
- Test results showing each filter's catch rate on known examples.
Common mistakes
- Filtering the base dataset and skipping data added later for fine-tuning.
- No secret scanning on scraped code, so live keys end up in the model and come back out in completions.
- Deleting removed items with no log, which leaves you unable to tell a working filter from a broken one.
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.