Home / Glossary / Allow, ask or deny
Glossary
Allow, ask or deny: how policy for AI agents works
Allow, ask or deny is a policy model in which every action an AI agent tries gets one of three verdicts. Allow lets the action run. Ask puts the action to the user as a question. Deny stops it. The model gives security teams a middle setting between letting an agent do anything and blocking it outright.
By the AIBound team · Updated 17 November 2026
Key takeaways
- Allow, ask or deny gives every action an AI agent tries one of three verdicts.
- Ask puts the action to the user as a question, so a person decides the cases a rule should not settle alone.
- The model works when the rule is checked before the action runs and the user has no way to switch it off.
Why AI agents need three verdicts
A rule with only two answers forces a bad choice. Block too much and employees stop using the agent or find a way around it. Allow too much and one wrong instruction can delete files or send data out.
The middle verdict fixes this. Many actions are fine most of the time and risky some of the time, such as pushing code or installing a package. Ask lets the agent keep working while a person makes the call on those.
How allow, ask or deny works
Each rule names a kind of action and the verdict it gets. The check happens before the action runs, so a denied command never starts.
- 1 An admin writes the rules, for example deny reading secrets files and ask before deleting a folder.
- 2 The agent tries an action, such as running a command, reading a file or calling a tool.
- 3 The action is matched against the rules and gets its verdict: allow, ask or deny.
- 4 Allowed actions run, asked ones wait for the user's answer, and denied ones are stopped.
Allow, ask or deny vs allow or block
Older controls, such as a firewall rule or an app blocklist, give two answers. Policy for agents adds a third, because an agent's actions vary far more than an app's. The same agent can list files one minute and try to remove them the next.
| Question | Allow, ask or deny | Allow or block |
|---|---|---|
| How many verdicts? | Three: allow, ask or deny | Two: allow or block |
| What is judged? | Each action the agent tries | A whole app, site or connection |
| Who decides the grey cases? | The user, when the rule says ask | The rule settles every case in advance |
Sources
- OWASP LLM06: Excessive Agency · genai.owasp.org
Sources checked 19 November 2026
Examples
Allow, ask or deny examples
FAQ
Frequently asked questions
How does allow, ask or deny policy for AI agents work?
Each rule names a kind of action and gives it a verdict. When the agent tries something, the action is checked against the rules before it runs. Allowed actions go ahead, asked ones are put to the user as a question, and denied ones are stopped.
What does "ask" mean in an AI agent policy?
Ask means the action is put to the user as a question before it runs. The agent pauses, shows what it wants to do and waits. If the user says yes, the action goes ahead. If the user says no, it stops. Ask suits actions that are sometimes fine and sometimes risky.
When should a rule deny instead of ask?
Use deny when no employee should be able to approve the action, such as reading credentials or sending data to an unknown host. Use ask when the right answer depends on context the user has. Too many questions teach people to click yes, so keep ask for real judgement calls.
Can employees change an allow, ask or deny rule?
That depends on where the rule lives. A setting in an agent's own local file can often be edited by the person using it. With AgentHarness from AIBound, the policy runs inside the agent, employees cannot switch it off, and an admin sets it for the company, a team or one device.