Topic: Identity

Your ISO 27001 controls already cover agents. Your evidence doesn't

ISO 27001 and SOC 2 already reach software actors. The evidence auditors sample never contains your agents. Five questions before your next audit.

In this post 8 sections
  1. The standard already counts software
  2. Why agents still pass
  3. Control by control
  4. ISO 42001 doesn’t fill the gap
  5. Where auditors are heading
  6. Five questions before your next surveillance visit
  7. Putting agents into the evidence
  8. Sources

You hold an ISO 27001 certificate, maybe a SOC 2 report as well. The access control section passed. Now there are agents in the business, connected to the CRM, the ticketing system and a cloud console, and you’re wondering whether the controls you trust still mean what you think.

I expected to write that agents satisfy the letter of these controls while defeating their purpose. That turned out to be wrong, and the real answer is more useful. The standards already reach agents. What doesn’t reach them is the evidence your audit samples.

The standard already counts software

ISO/IEC 27002, the guidance behind ISO 27001’s controls, defines its terms in clause 3. An entity is an “item relevant for the purpose of operation of a domain that has recognizably distinct existence”, and its examples include “a software application, a service or a website”. Authentication is “provision of assurance that a claimed characteristic of an entity is correct”. Non-repudiation is the “ability to prove the occurrence of a claimed event or action and its originating entities”.

SOC 2 says something similar. Under CC6.1, the points of focus say “persons, infrastructure and software are identified and authenticated prior to accessing information assets”, and that new software is “registered, authorized, and documented prior to being granted access credentials”.

So an agent is plainly in scope. A careful auditor could raise it today.

Why agents still pass

Audits test controls by sampling evidence, and the populations they sample from are built around people: the HR joiner and leaver lists, the quarterly user access review, the privileged account register, sign-in logs keyed to named users. That is my reading of how these audits are usually run, not a published finding, but it is the mechanism. An agent running on an engineer’s token, an old integration service account or an OAuth grant someone clicked through isn’t in any of those lists as a separate thing. Every sample comes back clean.

How common is that setup? In a Cloud Security Alliance survey commissioned by Aembit, 43 per cent of organisations relied on shared service accounts for agents, and 31 per cent let agents operate under a human’s identity. In Schellman’s 2026 survey of 525 US organisations, 74 per cent believed they could pass an AI compliance audit today, and 27 per cent described their governance as fully mature.

Here is how that plays out, control by control.

Control by control

5.15 Access control, and CC6.1. You trust this to mean access is authorised and restricted to business need. The policy exists, and the credential the agent uses was approved for someone. ISO 27002 defines access control as access that is “authorized and restricted based on business and information security requirements”. Nobody authorised the agent’s access or set its business need. Its reach is whatever the borrowed credential can do.

5.16 Identity management. You trust every actor to have a managed identity from creation to removal. The engineer and the service account each have a joiner record. The agent has no identity of its own, so it has no life cycle to manage. The joint “Careful adoption of agentic AI services” guidance, co-authored by Australia’s ASD with five other agencies, says developers “should construct each agent as a distinct principal”.

5.17 Authentication information. The token was issued properly to a named person or integration. It is now used by a second actor its issuer never registered. The same guidance warns that tokens get stolen “when organisations keep them static, share them across multiple agents, or protect them poorly”.

5.18 Access rights, and CC6.2 and CC6.3. The quarterly access review is complete and signed, and it is accurate about every human on it. It never asks what a second, non-human actor does with their access.

8.2 Privileged access rights. The privileged account register is short and every entry has an approver. But a user with modest rights can steer a privileged agent into doing what they couldn’t do themselves. The joint guidance calls this the confused deputy: “a low-privileged user manipulates a high-privileged agent to perform actions the low-privileged user couldn’t do directly”.

8.5 Secure authentication. The human signed in with MFA. The agent rides a token issued after that sign-in, and if its rights are checked only at start-up, the guidance notes “a malicious actor can exploit a stale ‘allow’ decision”.

8.15 Logging. The logs exist, are kept and are protected. The log line names the human or the service account, so it can’t prove the agent was the “originating entity”. The guidance says it plainly: acting under a trusted identity, “the system produces audit logs that appear legitimate and delay detection”. I covered what an agent record needs in identity is the control plane for AI agents.

8.16 Monitoring activities, and CC7.2. Detection is tuned to the account’s normal behaviour, so the agent’s activity blends in. In the CSA survey, 68 per cent couldn’t clearly tell agent activity from human activity.

ISO 42001 doesn’t fill the gap

If you’ve added ISO/IEC 42001, the AI management system standard, it helps less here than you might hope. None of its 38 Annex A control titles is about identity, authentication, credentials or access rights. Its event logging control, A.6.2.8, is about deciding at which life-cycle phases to record logs, not about who acted. You can hold both certificates and still run agents on shared credentials.

Where auditors are heading

I found no published audit finding, certification body guidance or AICPA criterion on agent access yet, so I won’t tell you auditors are raising nonconformities. The signals are there, though. KPMG describes changes to the Trust Services Criteria as forthcoming, with sessions on when AI is relevant to a SOC examination. A-LIGN’s Patrick Sullivan put the auditor’s view well: if an agent shouldn’t take an action, “it should lack the technical ability to take it, not merely the instruction”. And the joint guidance tells organisations to handle agents “within established cyber security frameworks rather than treating it as a separate or standalone discipline”. That is your existing ISMS.

Five questions before your next surveillance visit

Each one asks for evidence, not a policy statement. Put them to your own team first.

  1. Show me every agent that can act in our in-scope systems, and the identity each one uses. If the answer is a list of people and service accounts, agents have no identity life cycle.
  2. In last quarter’s access review, which lines were agents, and who signed them?
  3. Which agents can reach anything on our privileged access register, and through whose credentials?
  4. Pick one agent action from last month. Can our logs prove it was the agent, which tool it called, and which person it acted for?
  5. If one agent misbehaved today, could we cut off that agent alone without breaking the person or integration whose access it borrows?

If you hold a SOC 2 report, add a sixth: do any of our control descriptions say “human” where an agent now does the work? The least agency post covers how far an agent should be allowed to act before a person checks.

Putting agents into the evidence

The fix is to make agents show up in the same evidence the controls already ask for. Each agent gets its own identity, recorded with an owner, and only the access it was granted. Risky actions wait for a named person. Every call is recorded against the agent and the person it acted for, so the access review and the logs can sample it like anything else.

That is what AgentValet provides, and the trust page shows how the record is kept.

Sources

  • ISO/IEC 27002:2022, clause 3 terms and definitions (ISO Online Browsing Platform preview): iso.org
  • ISO/IEC 27001:2022 Annex A control titles, as mapped by ISMS.online: isms.online
  • AICPA 2017 Trust Services Criteria (revised points of focus 2022), CC6.1, as reproduced by Cloudaware: cloudaware.com
  • CISA, NSA, ASD’s ACSC and others, Careful adoption of agentic AI services (1 May 2026): cyber.gov.au
  • Cloud Security Alliance survey commissioned by Aembit (24 March 2026): cloudsecurityalliance.org
  • Schellman, 2026 State of AI Governance Report (29 July 2026): schellman.com
  • ISO/IEC 42001:2023 Annex A control index, as listed by TCSA: tcsa.in
  • A-LIGN, AI agents are running in your business (15 May 2026): a-lign.com
  • KPMG, The future of SOC reporting: mastering AI risks and the changing Trust Services Criteria (2026): kpmg.com

Filed under

  • identity
  • compliance
  • iso-27001
  • soc-2

Give your agent only the access it needs

Grant each agent a narrow set of actions, hold the risky ones for your approval, and keep a signed record of every call.