Risk · R-172 · Agentic & Autonomous Systems
Action on stale or withdrawn authority
An agent's action takes effect after the approval, delegation or condition that justified it has expired, been withdrawn or ceased to hold, so the effect occurs without valid authority.
- Severity
- High
- Audience
- Heads-of · Practitioner
- Frameworks
- Article 14 · ASI03
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-17
What it is
Authority for an agent's action is established at some point: a user's request, an approval, a delegated grant, a condition such as an account being in good standing. The action takes effect later, sometimes much later, after a queue, a retry, a hand-off or a different execution path. In between, the authority can change. The user withdraws the request, an approver's own authority lapses, the delegation expires, the customer's circumstances change. The risk is that the action proceeds on authority checked at the start and not at the point of effect. It differs from an agent acting beyond its permissions: here the permission existed and no longer applied. Commitments that were never within an agent's authority are covered in unauthorised commitments by agents.
How it happens
- Authority is checked when a task starts and not again before the action that matters.
- A failed step is retried hours later on the original approval.
- A queued task outlives the revocation of the credential that created it.
- An alternative execution path skips the check the main path performs.
- Revocation relies on a list that the executing component cannot reach.
Why it matters
Where the change of authority was the point, acting on the old position is the very harm the change was meant to prevent: a customer cancels, a payment is stopped, an approver is removed after a fraud. Every control that checked authority at the start will show that it operated.
Warning signs
- Approvals have no expiry and are not tied to the parameters approved.
- Retried and queued actions are not re-authorised.
- Revocation takes effect only for new tasks.
- No test changes authority partway through a task.
Who owns it
The business owner of the process decides which conditions must still hold when an action takes effect. Engineering builds the check into every execution path. Identity and access management sets credential lifetimes and revocation.
In practice
A customer asks an agent to move funds and, a minute later, cancels. The cancellation reaches the conversation, but the transfer had already been queued after a timeout, and the queue worker executes it on the original instruction. The same pattern generalises to any action that can outlive the reason for it.
Mitigating controls
Primary controls
- Authority check at the point of execution Preventive
- Checks authority, approval and conditions again at the point of effect. Immediately before an agent's action takes effect, its authority is confirmed again: the delegation, approval, credential and business conditions must all still hold, or the action does not proceed.
Supporting controls
- Agent action tracing Detective
- Shows which authority an action relied on when it took effect. Every agent action is recorded with the instruction, plan, tool call, result and identity that produced it, in a form that allows the sequence to be reconstructed.
- Approval gates for high-impact agent actions Preventive
- Binds an approval to the exact action, with an expiry. Before an agent takes an action above a defined impact threshold, or one that cannot be reversed, a designated person reviews the exact action and approves or rejects it.
- Least-privilege agent identity and permissions Preventive
- Short credential expiry means lapsed authority fails safe. Each agent has its own identity, with permissions scoped to its task and to the user it is acting for, short-lived and reviewed like any other entitlement.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| EU AI Act | Article 14 | Human oversight Human oversight requires that a decision to stop or override actually takes effect. |
| OWASP Top 10 for Agentic Applications | ASI03 | Identity and Privilege Abuse Authority that has lapsed is still being exercised. |
Sources
- EU AI Act: Regulation (EU) 2024/1689 European Union (EUR-Lex) · official text
How to cite
Axiom Verity (2026). Action on stale or withdrawn authority. Reference Library, R-172. https://axiom-verity.com/risks/action-on-stale-or-withdrawn-authority. Reviewed 2026-11-17.
Found an error or a better source? Report a correction.