Skip to content
Enterprise mobility
ASD ISMISM-2095

ASD ISM 2095 Bar personnel on private devices from granting access to unapproved AI agents

What it says

Personnel using privately owned mobile devices or desktop computers to access OFFICIAL: Sensitive or PROTECTED systems or data are disallowed from granting access to unapproved artificial intelligence agents.

Quoted as published. Everything else on this page is written by Control Stack.

In plain English

Personnel reaching OFFICIAL: Sensitive or PROTECTED systems or data from privately owned mobile devices or desktop computers must not be allowed to give unapproved AI agents access.

Personnel using privately owned mobile devices or desktop computers to access OFFICIAL: Sensitive or PROTECTED systems or data must be disallowed from granting access to unapproved artificial intelligence (AI) agents. An AI agent is software that uses an AI model to act for a person, such as reading mail or using other tools. Granting access includes signing an agent in, approving its connection or letting it read the screen. Only unapproved agents are covered, so the organisation decides which are approved. Disallowed sets a rule, not necessarily a technical block.

Why it matters

An agent given someone's access can read and act on whatever they can reach, and a grant made on a personal device can go unseen. Disallowing unapproved agents limits that access to agents the organisation has approved.

How to implement

  1. Decide which AI agents are approved.

    The control sets no approval criteria, so the organisation sets them, based on what data an agent can reach and where that data goes. Record approved agents in the AI agent register (ISM-2134), and name who approves new ones.

  2. State the rule for personnel using privately owned devices.

    Put it in the conditions of use and the general-purpose AI usage policy (ISM-2074). Spell out what is disallowed: signing in an unapproved agent, approving its connection, sharing a token or password, installing its browser extension, or giving it screen or file permissions.

  3. Block the grants a platform can block.

    Turn off user consent to third-party apps in the identity platform, so new connections need administrator approval. Restrict who can create API tokens. For a bought service, use its admin settings for connected apps and AI features.

  4. Find grants that got through.

    Compare connected apps, consent grants and API tokens on these systems with the approved list, at an interval you set (the control sets none). Revoke anything unapproved and follow up with the person concerned.

Keep it running

Recheck when an app, browser or operating system update adds an AI feature, because these can arrive on personal devices without any request. A newly popular agent, or a new system reachable from these devices, should trigger a review of connected apps.

Evidence and audit

  1. Check 1 of 3What an auditor asks for

    Ask for

    Connected-app, consent and token records for OFFICIAL: Sensitive or PROTECTED systems, compared with the approved agent list.

    Look at

    Whether any agent holds a consent, token or account without being on the list, and whether the records can be tied to people using privately owned devices.

    Good evidence

    Every agent holding a consent, token or account on those systems is on the approved list, and the search covered grants made from privately owned devices.

  2. Check 2 of 3What an auditor asks for

    Ask for

    A test with a test account on a privately owned device, trying to connect an agent that is not on the list to one of those systems.

    Look at

    Whether the connection is refused or held for approval, or whether it completes and is then found.

    Good evidence

    The connection is refused or held for administrator approval. Where the platform cannot block it, the grant appears in the connected-app records and is revoked, so a grant nobody detects fails.

  3. Check 3 of 3What an auditor asks for

    Ask for

    The rule as given to these personnel, and a conversation with a sample of them chosen from sign-in logs.

    Look at

    Whether the rule covers them and says what counts as granting access, and whether each person knows it and can say how to find the approved list.

    Good evidence

    The rule names what counts as granting access. Each sampled person had it before using the device for these systems, and can say how to check whether an agent is approved.

Mappings

Closest match in ISO/IEC 27001: A.5.10, Acceptable use of information and other associated assets, which this control partially meets.

ISO 27001

ISO/IEC 27001: 4 mappings

  • A.5.15Access controlPartially meets
  • A.5.18Access rightsPartially meets
  • A.8.3Information access restrictionPartially meets

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