Est.

ISO IEC 42001 AI Management System Requirements for Agent Programs

The standard needs reinterpretation to govern agents that act autonomously between human reviews.

· 10 min read · Updated
Cover illustration for “ISO IEC 42001 AI Management System Requirements for Agent Programs”
Adoption Frameworks · September 24, 2026 · 10 min read · 2,328 words

Its clauses (4 through 10) address organizational context, leadership, planning, support, operation, performance evaluation, and improvement, with Annex A providing a reference set of controls. It is voluntary, carried out by independent certification bodies, and it lasts three years with annual surveillance audits in between, and it has become a meaningful signal in enterprise procurement and supply chain risk reviews.

The standard was published in 2023, before agentic AI reached production at any real scale, and it was built to manage risk to the organization deploying the AI, not risk produced by autonomous systems acting on that organization's behalf out in the world. That timing means the standard governs posture, the policies, assessments, and controls an organization puts in place around an AI system, not runtime, which is the gap it leaves. Runtime is different: it is the moment-to-moment behavior of a system acting across tools, credentials, and external services with less human involvement at each step. Every clause in the standard still applies to agentic deployments. What changes is how much interpretation each clause needs before it actually closes the gap it was meant to close.

Agentic AI and the clause interpretation gap

Agentic AI is not a scaled-up version of the AI systems ISO 42001 implicitly assumes. It plans and breaks tasks into steps, invokes external tools on its own through APIs and MCP servers, runs multi-step action chains without approval at each step, changes external state by sending emails, executing transactions, or writing files, and adjusts its next move based on what the last step returned. The European Data Protection Supervisor and the European Commission's StepUp Startups initiative Agentic AI report both describe these properties as the defining differentiators that separate an agent from a standalone model, not as unusual edge cases that show up occasionally.

Each of those properties creates its own control gap, because the system was built to manage AI that humans run directly. Autonomous tool invocation means the set of tools an agent can reach is resolved dynamically, at the moment the agent decides to act, outside any static access list a security team wrote in advance. Every new tool integration needs its own authentication, so credential delegation multiplies OAuth tokens, API keys, and service account permissions faster than any human identity and access management policy was built to track. Multi-step execution without per-step approval means a human can review what happened before a chain starts or after it finishes, but not in between, while the chain is running. So when an agent makes a mistake, or a compromised agent acts, the consequences land in the real world before any log gets reviewed. Feedback-driven adaptation means an agent's behavior can shift within a single session.

Each of these gaps lines up with a clause in ISO 42001 that partially covers it already. None of them require throwing the standard out. All of them require reading the clause with agentic behavior in mind rather than the human-operated systems the clause was originally written against.

Why the MCP threat landscape makes clause interpretation urgent

The control gaps described above are not a future risk that compliance teams can schedule for next year's planning cycle. They are already being exploited in documented production incidents.

On January 20, 2026, Cyata security researcher Yarden Porat published an exploit chain against Anthropic's own official Git MCP server, built on three separate CVEs: a path traversal flaw (CVE-2025-68143), an argument injection flaw (CVE-2025-68144), and a second path traversal and repository restriction bypass (CVE-2025-68145). The target here was not some obscure community tool. It was a server published by the company that originated the Model Context Protocol itself. Even vendor-supplied MCP servers carry attack surfaces nobody reviewed closely enough. That single case exposes the exact gap Clause 6 and Annex A are meant to close: an organization cannot assume a tool is safe because of who published it.

Supply chain poisoning tells a parallel story at larger scale. Antiy CERT confirmed at least 1,184 malicious skills circulating in the ClawHub agent-skill ecosystem, and Snyk's ToxicSkills audit found that a meaningful share of the skills it reviewed carried at least one critical security issue. That is a supplier management failure playing out across an entire ecosystem, which is exactly the kind of gap Annex A's supplier controls were written to prevent, just never written with an agent-skill marketplace in mind.

What makes all of this urgent rather than theoretical is how poorly most organizations can currently see any of it happening. These incidents are occurring in environments where ISO 42001's monitoring and audit clauses are nominally in place and practically unimplemented for agents. That is the condition the rest of this piece addresses, clause by clause.

Clause 6 planning and risk assessment when the agent's action space is not knowable at design time

Clause 6 requires you to identify AI-related risks and plan the controls that manage them. In a traditional AI deployment, that works because the system's capabilities are fixed at design time. In an agentic deployment, the full action space cannot be enumerated at design time, because it gets resolved dynamically against whatever MCP servers and tools happen to be available when the agent runs.

Risk assessment under Clause 6 has to shift from cataloguing specific tools to bounding categories of permissible access: what kinds of tools an agent may invoke, under what conditions, with what credential scope. That is a different exercise than listing out named integrations one by one, and it has to stay current as the agent's environment changes.

The practical version of this looks like a live inventory of MCP servers in scope for each agent deployment, with any addition to that inventory treated as a change event that triggers a fresh risk assessment, and a default-deny posture for any tool category that was never explicitly approved. Clause 6 also requires the AI impact assessment to cover more than the agent's intended function. It has to account for unintended environmental state modification: sending emails, executing transactions, or modifying records that the agent was never meant to touch, because those are the consequences that land before any human reviews a log.

There is an obvious objection here. Organizations running hundreds of agents cannot realistically run a full risk assessment every time a new MCP server gets registered. Tiered assessment assigns a baseline risk profile by agent role, triggers a lightweight re-assessment only for meaningful scope changes, such as a new tool category, an elevated credential scope, or a new data source, and automatically flags whenever an agent calls a tool outside the profile it was assessed against.

Annex A accountability and supplier management for MCP servers as third-party components

Annex A.10.2, covering allocation of responsibility, and A.10.3, covering suppliers, together require that any externally supplied component an agent depends on has a named owner and a supplier review attached to it. An MCP server supplied by a third party, or pulled from a community registry, meets that definition exactly, whether the organization has been treating it that way or not.

The interpretation this calls for is direct: every MCP server an agent is permitted to invoke, commercial, open-source, or community-published, needs to be treated as a supplier relationship under A.10.3, with a review covering authentication posture, update cadence, known vulnerabilities, and data handling practices. Under A.3.2, accountability for the agent extends past the agent itself to the external models, MCP servers, and data sources it relies on. So the accountability chain does not stop at the agent's own API boundary.

The ClawHub case shows what happens when that chain gets treated as inert infrastructure instead of a set of supplier relationships: 1,184 malicious skills reached production agent deployments with no supplier-style review standing in the way. The practical fix is a registry of approved MCP servers functioning like a vendor allowlist, where every entry carries a named owner, a review date, and a defined trigger for re-review, such as a new CVE, a major version change, or a transfer of ownership. The Anthropic Git MCP Server exploit chain reinforces why this applies even to vendor-supplied servers. If a major LLM provider publishes a server, supplier trust in that provider does not transfer to it automatically.

Once an organization knows which MCP servers are approved, a second question follows immediately: what credential does each agent use to reach them. That question belongs to Clause 8.

Clause 8 operation controls for credential delegation and non-human identity management

Clause 8 governs operational planning and control, and the sharpest operational gap in agentic deployments is that agents sit outside the identity and access management infrastructure organizations already built for human users. That infrastructure was designed around people logging in with a single identity. Agents don't work that way.

Clause 8 needs to explicitly cover agent identity lifecycle management: provisioning, scoping, rotation, and revocation of the credentials issued to agents, treated as a category of its own rather than folded into human user accounts or legacy service accounts. A single agent interaction can require OAuth tokens from a cloud provider's IAM endpoint, a separate OAuth flow through an MCP authorization server for tool access, and vendor-specific API keys for the LLM provider itself, each issued by a different authority, each carrying different scopes and expiry rules. Clause 8 controls have to cover that full set of credentials, including the agent's primary identity.

Hardcoded secrets spreading across agent configurations, environment variables, and orchestration frameworks are a Clause 8 operational control failure in their own right. The credentials sit outside the organization's credential management processes entirely. The Drift and Salesforce OAuth token theft case shows what that failure costs in practice: when an agent authenticates on a user's behalf, a flaw anywhere in that credential flow exposes the user's own authorization credentials, not just the agent's.

The governing principle is least-privilege, just-in-time credential issuance, scoped to each agent identity and each specific task rather than granted broadly up front. That principle does more than tighten access control. A credential issued for one task, used once, and retired leaves a far cleaner trail than a standing credential an agent reuses indefinitely, making per-action audit logging possible.

Clause 8.4 and Annex B.4 event logging requirements for tool-call chains

Clause 8.4 covers perception and event logging, and Annex B.4 covers human oversight logs. Together they require that an organization can show what an AI system did and when. In a multi-agent deployment running across multiple MCP servers, the events that would need to add up to a complete audit record are scattered across systems that were never built to produce logs that correlate with each other.

A complete tool-call record has to include, at minimum, the identity that initiated the call (user or agent), the MCP server and specific tool invoked, the arguments passed, the target system reached, the response received, and the timestamp, and that record has to exist for every step in a chain, not just the request that started it. So anything less leaves gaps exactly where an incident investigation would need detail most.

The Cisco agent-to-agent lateral movement case shows what those gaps cost. The path an attacker took from a compromised low-privilege data query agent to a high-privilege financial approval agent was only reconstructable after the fact, because the interface the two agents used to communicate never produced a correlated log tying their sessions together. That is not an unusual failure mode. A Gravitee survey found that only 24.4% of enterprises have full visibility into which AI agents are communicating with each other, which makes agent-to-agent communication the most common source of audit gaps in practice, well ahead of external tool calls.

To close that gap, you need distributed tracing, built on something like OpenTelemetry, applied at the tool-call level with trace IDs that persist across agent handoffs and across MCP server boundaries. General application logging will not satisfy what Clause 8.4 and Annex B.4 are actually asking for. Only a tracing architecture built to follow an action across every agent and server it touches can produce the record the standard assumes exists.

Extending Clause 9 performance evaluation to detect runtime intent drift

That cadence assumes the thing you are measuring changes slowly, at the pace of model versions and retraining cycles.

Agentic systems break that assumption. An agent can be adapting its behavior within a single session, responding to intermediate results in ways that shift what it does next, long before the next scheduled evaluation window arrives. A performance review conducted monthly, or even weekly, is structurally incapable of catching a drift that happens over the course of one multi-step task. An agent can start a session performing exactly the function it was deployed for and end that same session doing something subtly different, and a post-deployment evaluation schedule would not catch the shift until whatever consequence it produced had already landed.

Extending this to cover runtime intent drift means treating it as its own category of thing to measure, separate from model accuracy or task success rate. That requires monitoring that operates on the same timescale as the agent's own action chains. It means comparing what an agent actually did, tool by tool, against what its assessed risk profile under Clause 6 said it was authorized to do, and flagging the moment those two things diverge rather than waiting for a quarterly review to notice, since the correlated, tool-call-level tracing required to demonstrate what an AI system did and when under Clause 8.4 and Annex B.4 is also the raw material needed here: a trace that shows a credentialed agent reaching for a tool category outside its baseline profile is an intent-drift signal in real time, not just a forensic artifact after the fact.

Performance evaluation under ISO 42001 was built to answer whether a model still performs the way it did when it was validated. For agentic systems, you have to ask that question continuously, against the agent's actual runtime behavior, for the standard's intent to hold.

Sources

  1. AI Agents Under EU Law
  2. Underwriting the Agent Economy: The Blueprint for an AI Insurance Stack

More in Adoption Frameworks