Verify the Source and Integrity of AI Training Data
Check where the data used to train artificial intelligence (AI) models comes from and confirm it has not been tampered with before it is used.
Plain language
Artificial intelligence (AI) models learn from large sets of data, and the quality of that data shapes how the model behaves. This control means you must know exactly where your training data came from (its source) and confirm it has not been changed, corrupted, or poisoned along the way (its integrity). If you train a model on data of unknown origin or data that someone has secretly altered, the model can make wrong, biased, or unsafe decisions, and you may not realise why.
Framework
ASD Information Security Manual (ISM)
Control effect
Preventative
Classifications
NC, OS, P, S, TS
ISM last updated
Dec 2025
Control Stack last updated
18 June 2026
E8 maturity levels
N/A
Guideline
Guidelines for software developmentOfficial control statement
The source and integrity of training data for AI models is verified.
Why it matters
If training data of unknown origin or that has been secretly altered is used, the AI model can produce wrong, biased, or manipulated outputs that the organisation trusts and acts on without realising the cause.
Operational notes
Re-verify source and integrity every time a dataset is refreshed or re-supplied, not just on first use, and keep provenance and checksum records for as long as the model is in service.
Implementation tips
- The data team should record where every training dataset comes from (the supplier, the internal system, or the public source) and keep this provenance record alongside each dataset so its origin can always be traced back.
- An engineer should generate a checksum (a unique digital fingerprint, such as a SHA-256 hash) for each dataset when it is received and recheck that fingerprint before the data is used for training, so any change to the file is detected.
- The procurement or vendor manager should require data and AI suppliers to provide written confirmation of where their data originated and that it has not been altered, and capture this in the supply contract.
- The AI project lead should run basic quality and sanity checks on incoming data (looking for unexpected entries, duplicated records, or signs of deliberate poisoning) and document the checks before approving the data for training.
- The IT manager should store approved training datasets in a controlled location with restricted access and version control, so only verified data is used and any later change is logged and reviewable.
Audit / evidence tips
- Askthe provenance records for a sample of training datasetsLook ata clear, named source for each one (vendor, internal system, or public repository)Goodtraces every dataset back to a documented, legitimate origin with no gaps
- Askhow the organisation confirms training data has not been alteredLook atchecksums or hash values recorded on receipt and re-verified before useGoodshows the fingerprint was recorded and matched, with evidence of what happens when it does not match
- Askto see the contracts or agreements with data and AI suppliersLook atclauses requiring the supplier to confirm data source and integrityGoodshows these obligations are written into the agreement, not just assumed
- Askevidence of the quality and integrity checks run on incoming dataLook atdocumented sanity checks and the results, including any datasets that were rejectedGoodshows checks are routine and that bad data is actually turned away
- Askwho can access and change the approved training datasetsLook atrestricted access, version control, and change logsGoodshows only authorised staff can touch the data and every change is recorded
Cross-framework mappings
How ISM-2087 relates to controls across ISO/IEC 27001, ISO/IEC 42001, Essential Eight, and ASD ISM.
ISO 27001
| Control | Notes | Details |
|---|---|---|
sync_altPartially overlaps(2)expand_less | ||
| Annex A 5.21 | ISM-2087 requires the organisation to verify the source and integrity of training data used for AI models to prevent data poisoning | |
| Annex A 8.30 | ISM-2087 requires the organisation to verify the source and integrity of training data used to build AI models | |
ISO 42001
| Control | Notes | Details |
|---|---|---|
sync_altPartially overlaps(3)expand_less | ||
| Annex A 4.3 | Annex A 4.3 requires documenting the data resources utilised for the AI system | |
| Annex A 6.2.7 | Annex A 6.2.7 requires the organisation to determine and provide AI system technical documentation appropriate to different interested pa... | |
| Annex A 7.5 | Annex A 7.5 requires a documented, end-to-end process for recording the provenance of data used in AI systems across the full life cycle | |
These mappings show relationships between controls across frameworks. They do not imply full equivalence or certification.
Related ASD ISM controls in Software development
See all Guidelines for software development controls, or browse the full ASD ISM library.