This website uses cookies

Read our Privacy policy and Terms of use for more information.


A signed approval does not govern an AI agent. It governs the day it was signed.

On 11 November Cequence Security, with input from KPMG in Singapore, published a report on governing agents as they act. In a controlled study on a Cequence research prototype, agents given only logins and fixed permissions tried to act outside their assigned role in 44 per cent of runs. When each agent was given a defined job and every action was checked against it, the figure fell to 1 per cent, and task completion rose from 76 to 95 per cent. The release does not say how many runs were made.

The approval was true on the day it was signed

Most firms govern an agent the way they govern a new supplier. It is reviewed, approved, recorded, and revisited at the next cycle. That works when the thing approved behaves after the signature as it did before. An agent does not. It chooses sequences of actions inside the permissions it was given, and each sequence is new.

Permissions say what an agent can enter. They do not say what it is for. An agent allowed to read customer records and issue refunds can do both in one run, in an order nobody reviewed, for a reason nobody wrote down. The approval was accurate on the day it was signed. It cannot be accurate about actions that did not yet exist.

A permission describes the door. It does not describe the errand.

Stale approval is the default

Call it stale approval: a sign-off from one moment, treated as covering every later action an agent takes. Last week's problem was unsigned reach, access that nobody decided. This week's is its mirror. Somebody did sign, once, and the signature has been read as current ever since.

Cequence's chief technology officer, Shreyans Mehta, put it in the release in one line: "An approval at step one shouldn't carry into step five."

The survey behind the report points the same way, with a caution. The report page cites 202 IT and security leaders in North America and in Europe, the Middle East and Africa, surveyed by Enterprise Management Associates for Cequence. Of them, 65 per cent had seen an AI agent act outside its intended scope. Cequence sells the control the report recommends and commissioned the survey. Read both as a vendor's evidence: enough to ask the question, too thin to set a figure for your own firm.

The regulator is asking a version of the same question. The UK data protection regulator, the ICO, opened a call for evidence on agentic AI on 8 October. It closes at the end of Friday 20 November, and the ICO has said the answers will shape its guidance and a statutory code on AI and automated decisions. As summarised by law firms, it asks how safeguards are implemented when agents act. Safeguards run at the point of action. An approval runs on the day it was given.

A design was approved. The behaviour never was.

Aviation re-issues the clearance at every phase

Aviation settled this long ago. A pilot cleared to taxi is not cleared to take off. Each phase of a flight carries its own clearance, read back and recorded, because conditions change between phases and permission for one says nothing about the next.

An agent working through a multi-step task is a flight with many phases. The refund approved at the start is the taxi clearance. A change to a record, a payment, a message to a customer: each needs its own. The report's recommendation is the same idea in a vendor's terms, a control outside the agent that approves, escalates or blocks each action as it happens. A firm does not need that vendor, or any vendor, to ask for the principle. Each consequential action is checked against the job the agent was given, at the moment it is taken.

Cleared to taxi is not cleared to take off.

What a board should hold

Four things, for each agent that can reach money, customer records or accounts.

The job. One plain sentence, signed by a named person and dated, saying what the agent is for. Not what it may access.

The limits that follow, with their values: an amount, a record type, a customer group.

What is checked when the agent acts, by something outside the agent, and who decides when an action is escalated.

The exceptions. Last month, how many actions fell outside the job, how many were blocked or escalated, and how many tasks were completed. The report recommends measuring completed work as well as blocks, and a board should want both, because a control that blocks everything fails as surely as one that blocks nothing.

None of this needs a particular product. If the exception count cannot be produced, that is the finding.

An approval is a date. Governance is everything after it.

Review the job whenever the agent's task changes, and read the exception record every month. If your board asked tomorrow for last month's exceptions for its widest-reaching agent, would the answer be a number or a policy?

Roche-Review. Executive Governance, with Human Observability and Adaptive Architecture.

Sources: Cequence Security, with input from KPMG in Singapore, Agentic AI Runtime Security, Governance, and Control (release of 11 November 2026, and the report page titled Delegating Outcomes, Not Authority; the full report was not read, so figures are as reported); ICO, call for evidence on agentic AI (opened 8 October, closes 20 November 2026), as summarised by TLT, Freshfields and LexisNexis.

The Roche-Review is the weekly publication of Dr Ivan Roche FRSS FRSA MInstP, Founder of Otopoetic Limited. Subscribe at roche-review.com.

/