ASD ISM 2086 Verify integrity of AI models, structures, and weights
What it says
The source and integrity of AI models, structures and weights are verified.
Quoted as published. Everything else on this page is written by Control Stack.
What it means
In plain EnglishEnsure AI models are genuine and have not been tampered with to maintain trustworthiness.
This control is about making sure that the AI models you're using are what you think they are. If someone changes them without you knowing, it could lead to wrong decisions or security breaches.
Free. Save your status, tasks and notes.
Why it matters
If AI models are tampered with, it can lead to wrong decisions which might harm your business financially or reputationally.
How to implement
The IT team should keep a list of all AI models currently in use.
Write down the source, version, and last updated date for each model to track its authenticity.
System owners should regularly check these AI models for any unauthorised alterations.
Use a tool that compares the current version of the model with prior, verified versions.
Procurement managers need to ensure that any new AI models come from trusted vendors.
This can be done by evaluating the vendor's reputation and security credentials before purchase.
Data scientists should document the structure and weights of each AI model they develop.
This involves creating a report that captures its key characteristics and how it was trained.
Managers should organise training sessions for staff to understand the importance of verifying AI model integrity.
Include real-life examples where tampered models led to business issues.
Keep it running
Regularly verify AI models to ensure they haven't been altered. Keep reports and logs updated for any checks and changes made.
Evidence and audit
- Check 1 of 5What an auditor asks for
Ask for
a list of all AI models in use: Request a database or document listing each model, its source, and its current version.
Look at
whether these details match with procurement records.
Good evidence
Consistent records across systems indicating careful tracking.
- Check 2 of 5What an auditor asks for
Ask for
evidence of regular integrity checks on AI models: Request the logs or reports from recent verification activities.
Look at
entries that confirm the model's structure and weights were checked for unauthorised changes.
Good evidence
Recent and detailed logs showing checks were conducted methodically.
- Check 3 of 5What an auditor asks for
Ask for
to see vendor evaluation records for AI models: Request past assessments or vendor reviews before any purchases were made. Look to see if the vendor's security practices and track record were considered.
Good evidence
A detailed evaluation process document showcasing informed decision-making.
- Check 4 of 5What an auditor asks for
Ask for
documentation on AI model training and development: Request detailed reports from data scientists about how each model was built.
Look at
specifics like algorithms used, dataset descriptions, and version history.
Good evidence
Clear and comprehensive reports that explain model development.
- Check 5 of 5What an auditor asks for
Ask for
training session records: Request attendance lists and session content from staff training events about AI model integrity.
Look at
the materials used and if they cover verifying integrity effectively.
Good evidence
Documented evidence of training with well-explained objectives and examples.
Mappings
Closest match in ISO/IEC 27001: A.5.21, Managing information security in the ICT supply chain, which this control partially meets.
ISO/IEC 27001: 1 mapping
This control partially meets A.5.21.
ISM-2086 requires that AI model source and integrity are verified to prevent unauthorised or malicious model substitution and tampering. Annex A 5.21 addresses ICT supply chain risk management processes, which supports ISM-2086 by establishing requirements for trusted model suppliers, secure acquisition channels, and assurance activities for third-party model components. Annex A 5.21 is broader and process-oriented for all ICT products and services, while ISM-2086 is a specific assurance requirement for AI model artefacts.
Mappings show how controls relate across frameworks. They do not imply full equivalence or certification.