The Compliance Cliff: Your Agents Aren't in Any Control Catalog, and August Is Coming

Search for the word "agent" in your SOC 2 control matrix. Do the same for your ISO 27001 statement of applicability and your PCI DSS documentation. In almost every enterprise I've been a part of, the search yields nothing. The compliance frameworks that apply to your security program were written for people, and later for machines. They do not consider an autonomous agent that can authenticate with valid credentials, operate at machine speed across different systems, and alter its own intent mid-session. Control catalogs describe two types of subject, people and machines, while an agent is a third type. A catalog that does not have a subject category for an agent cannot control that agent, regardless of the framework you use and the deadline you face.

That gap was a theoretical problem last year. It becomes a concrete one on August 2, 2026, when the EU AI Act's Article 14 human-oversight obligations take effect. And it becomes an operational one the next time an auditor comes into your environment, sees agents running in production, and asks which controls apply to them.

The Gap Is Real and It's Documented

This is not guesswork. A CISO undergoing an audit at present is facing an auditor who will first have to determine whether there are agents in the environment at all, and then discovers the word "agents" absent in the organization's policies. It is confirmed by practitioners: the compliance frameworks companies actually get audited against, SOC 2, ISO 27001, PCI DSS, do not operationalize agent identities. Agent identities do not appear in any control catalog most practitioners have encountered. I faced the same situation from the standards side: while reviewing the Coalition for Secure AI's Shared Responsibility Framework prior to publication, the gap I kept flagging was evidence, because a framework can allocate responsibility for agentic systems, but the evidence an auditor relies on to confirm that accountability exists is absent from most control sets.

Standards bodies have started to respond. Both the Cloud Security Alliance and NIST have begun work on agentic-AI profiles, proposing autonomy-tier classification and runtime behavioral metrics. That is the leading edge of the catch-up. However, a proposed profile from a standards body is not equivalent to a control that will be in the audit framework your assessor will use. The audit catalogs have not caught up, and they will not have caught up by August.

Why This Is Worse Than a Normal Compliance Gap

The absence of a control for a certain known type of item is a common example of a compliance gap. You either have a control for a firewall or not; there is a control in the framework for that. Agents are not like that; the framework does not provide a category for them. This creates three distinct problems.

The controls are misaligned. When an auditor asks, "which control governs this agent's access to cardholder data," there isn't an answer because the control in question is for people, or service accounts, and the agent falls under neither category.

There is little evidence. Even when you can assert a control conceptually applies, the evidence it calls for (audit logs that show who did what) typically cannot separate agent actions from human actions. If the logs were built around prompts and responses, then you will not have execution traces at the granularity of individual tool calls.

Ownership is undefined. Compliance frameworks expect that every system and every access point has an owner who is responsible. In many cases, agents do not have an assigned owner. An agent with no named owner is a control finding waiting to be written up.

What August 2 Actually Requires

According to Article 14 of the EU AI Act, high-risk AI systems must have meaningful human oversight. For businesses operating agents this means translating the oversight into two functions most organizations do not have: named accountability for agent deployments, and an execution trace that can reconstruct an agent's actions. Programs without both will face enforcement actions in addition to operational risks.

If you operate in the EU or serve EU users, some of your agents are in scope. By August 2 the capability has to already exist and be demonstrable.

What to Build Before the Auditor Shows Up

The positive aspect is that the compliance work is the same as the security work. Closing the audit gap requires exactly the controls that protect you operationally. Here's the order.

Start by compiling the agent inventory. You cannot define controls for agents that you have not enumerated. Create a record for each agent in production including what it can access, what identity it possesses, and who its owner is. This one document addresses the auditor's first three questions and is a prerequisite for the rest of the items.

Ownership is the anchor everything else hangs from. Every agent must be assigned a named human owner, as is the case with all systems in your CMDB. This resolves the ownership gap that Article 14 and every audit framework presume has already been addressed.

Map your existing controls to agent identities. For each control in your current control catalog that relates to identity, access, or logging, determine the application to agents. For controls that are not a direct fit, state the compensating control. Do it before an auditor does it for you, or the gaps turn into findings.

Build execution-trace capability. Implement logging at the tool-call level to reconstruct agent actions during an incident. This fulfills the oversight obligations of Article 14 and matches the traceability the Five Eyes agencies called for in their April 30, 2026 agentic-AI guidance. This will also be a requirement from your auditor.

Not every agent needs the same scrutiny. Borrowing the autonomy-tier approach from the CSA's agentic profile work, document the autonomy tier of each agent, from information-only to fully autonomous cross-domain action. The higher the level of autonomy, the more oversight the framework and regulator will require you to show.

The Part Nobody Wants to Hear

You cannot patch or purchase a solution to this problem. The gap is not a missing tool. It is that your entire compliance posture was built for a world of human users and static service accounts. Agents are a third category that the frameworks have not captured.

The frameworks will catch up. SOC 2, ISO 27001, and PCI DSS will operationalize agent identities and the audit catalogs will follow. But "eventually" is post-August 2 and it's post your next audit cycle. Enterprises that see value in getting agent governance right, including inventory, ownership, control mapping, and execution traces, will come prepared with the evidence. Those waiting for the frameworks to tell them what to do will have agents in production and nothing mapped to them.

The auditor will track down the agents. The only question is whether you found them first.