Skip to content
Software development
ASD ISMISM-2086

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.

In plain English

Verify where each AI model, its structure and its weights came from, and that none has been altered since it left that source.

The source and integrity of AI models, their structures and their weights must be verified. The ISM applies this to models you build or use in applications you develop. Source means where each came from, such as the publisher that released the model or your own training run. Integrity means none of them has been changed since, except by someone authorised. The structure is the model's architecture, usually defined in configuration files or code, and the weights are the values learned in training. Training data is a separate requirement (ISM-2087).

Why it matters

A poisoned model can hide a backdoor that gives attacker-chosen output when a trigger appears, yet pass normal testing. Checking source and integrity confirms the model is the genuine release, or what your own training produced, and not a look-alike or tampered copy.

How to implement

  1. List every AI model you build or use in applications you develop.

    Include base models beneath fine-tuned ones. For each, record the publisher, where it was obtained, the exact version, and every file that holds its structure and weights, such as configuration and weight files.

  2. Obtain each model from its publisher's official channel.

    Confirm the publisher's identity, such as its verified organisation account on a model hub or its own website, and watch for look-alike names. Pin the exact version or revision rather than taking the latest, so a changed model cannot arrive unchecked.

  3. Check every file against a trusted value.

    Verify the publisher's digital signature, or compare each file's cryptographic hash with one obtained over a secure channel, as ISM-2027 requires for software artefacts. For a model you build, record hashes as soon as it is produced and store them apart from the model, with separate write access.

  4. Recheck before each deployment.

    Make the deployment pipeline recompute the hashes of the structure and weight files, compare them with the trusted values and stop on any mismatch, so a copy altered in storage is caught before it is loaded. Close any manual route that places model files on servers without this check.

  5. Handle hosted models whose files you never receive.

    Verify the source by connecting only to the provider's authenticated endpoint and pinning a specific model version where offered. For integrity, obtain the provider's evidence of how it protects that model. Without such evidence the model is unverified, and the requirement is not met for it.

Keep it running

Repeat the check whenever a model's files change, such as a new version, fine-tuning or conversion to another format, recording new trusted hashes for what you produced. A publisher's security notice or a revoked signing key should prompt a recheck.

Evidence and audit

  1. Check 1 of 3What an auditor asks for

    Ask for

    The list of AI models with the recorded source and verification for each, and the model files actually found in registries, inference servers and deployment configuration.

    Look at

    Whether every model found in use is listed, whether each named publisher is the genuine one rather than a look-alike, and whether the recorded check covers the structure and weight files.

    Good evidence

    Every model found in use is listed with its genuine publisher or the internal build that produced it; each downloaded or built model is also listed with its exact version and a passed check of every structure and weight file, and each hosted model with its pinned version where the provider offers one, and the provider's integrity evidence.

  2. Check 2 of 3What an auditor asks for

    Ask for

    Fresh hashes, taken during the audit, of the structure and weight files of a sampled production model, with the trusted values recorded for them.

    Look at

    Whether each hash matches the trusted value in the organisation's verification record and, for a published model, whether that recorded value matches the publisher's signed or published value.

    Good evidence

    Every file of the sampled model, configuration as well as weights, matches the trusted value in the organisation's verification record, and, for a published model, that recorded value matches the publisher's signed or published value.

  3. Check 3 of 3What an auditor asks for

    Ask for

    A demonstration, in a test environment, of deploying a model with one byte altered in a weight file, then again with one byte altered in its configuration file.

    Look at

    Whether the pipeline refuses each altered model, and who can place model files on inference servers by any other route.

    Good evidence

    The pipeline stops each altered model before it is loaded, and nobody can place model files on inference servers by a route that skips the check.

Mappings

Closest match in ISO/IEC 27001: A.5.21, Managing information security in the information and communication technology (ICT) supply chain, which this control partially meets.

ISO 27001

ISO/IEC 27001: 1 mapping

Mappings show how controls relate across frameworks. They do not imply full equivalence or certification.