Skip to content
Software development
ASD ISMISM-1923Rescinded

ASD ISM 1923 Mitigate the OWASP Top 10 for large language model applications

Rescinded in the December 2025 ISM

ISM-1923 is no longer part of the ISM. This page keeps it for reference, for organisations still working to an earlier version.

The existing control recommending that the OWASP Top 10 for Large Language Model Applications are mitigated in the development of large language model applications was rescinded.
ASD, ISM December 2025 changes

What it said

The OWASP Top 10 for Large Language Model Applications are mitigated in the development of large language model applications.

Quoted as published. The guidance below is written by Control Stack.

In plain English

Under ISM editions that included it, large language model applications had to be developed with each risk in the OWASP Top 10 for Large Language Model Applications mitigated.

The control required the risks in the OWASP Top 10 for Large Language Model Applications to be mitigated when developing large language model (LLM) applications, software that uses an AI model that generates text. The list covers ten risks, such as prompt injection and excessive agency. Mitigated means each is reduced by measures taken during development. The statement names no version of the list. No current ISM control names the list, though several address its risks separately, such as ISM-1924 on adversarial prompts.

Why it matters

LLM applications take instructions in plain language, so a crafted prompt, or text hidden in a document the model reads, can make it leak data or misuse the tools it can reach. Mitigating each listed risk during development builds defences in before release.

How to implement

  1. List the LLM applications being developed.

    Include in-house builds, contractor builds for the organisation and features in existing software that call a model through an API. Compare the list with code repositories and model provider accounts so none is missed, and record which version of the list each is assessed against.

  2. Work through all ten risks for each application.

    For each risk, record whether it applies to this design, the measures that mitigate it and where they sit in the code, configuration or development records. A 'not applicable' entry needs a reason that holds for this design. For a contractor's build, require the record and its test results as deliverables.

  3. Build the mitigations into the application.

    Current ISM controls describe several, such as evaluating prompts for adversarial inputs (ISM-1924), content filtering for sensitive data exposure (ISM-2094), fine-grained permissions (ISM-2092) and rate limiting of inference queries (ISM-2090). Enforce them outside the model where you can, because a manipulated model can ignore its instructions.

  4. Test each mitigation before release.

    Add test cases that attempt each applicable risk, such as instructions hidden in a document the application reads, and fix failures before release. Re-run them when the model, prompts, tools or data sources change.

Keep it running

Re-assess an application when its model, prompts, tools or data sources change, and when OWASP publishes a new version of the list, since risks can be added or redefined. Keep each risk record with the application's design documents.

Evidence and audit

  1. Check 1 of 3What an auditor asks for

    Ask for

    The list of LLM applications developed by or for the organisation, compared with code repositories and model provider accounts or billing.

    Look at

    Whether every application that calls a large language model is listed, and whether each has a risk record.

    Good evidence

    Every application found that calls a large language model is listed, and each has a record covering all ten risks in the version of the list it was assessed against.

  2. Check 2 of 3What an auditor asks for

    Ask for

    A sampled application's risk record, traced to its design, code, configuration, development records and test results.

    Look at

    Whether each recorded mitigation exists, whether each one in the build or configuration was tested, and whether each 'not applicable' reason holds.

    Good evidence

    Each of the ten risks is mitigated by a measure present in the build or its configuration, or by a development activity its records show was carried out, or genuinely does not apply to this design; every measure in the build or configuration has a passing test, and no risk is merely accepted or deferred.

  3. Check 3 of 3What an auditor asks for

    Ask for

    An attack test against a sampled application in a test environment, such as instructions hidden in a document it processes or a request for a tool action outside its task.

    Look at

    Whether the recorded mitigations hold.

    Good evidence

    The hidden instructions cause no data disclosure, no action outside the application's task and no false or planted content in the output, and the out-of-scope tool action is not carried out, matching what its risk record claims.