Your AI agent is approved. Is its next action?

AI agents can attempt actions nobody intended, even after they’re approved to run. 

 

This post explains how Airlock Digital Agentic Control checks each command a supported agent proposes before it runs, and how Agentic Steering helps the agent find a permitted way to finish the work based on your policy.

Approving an AI agent to do a job is not the same as approving every action it chooses along the way.

AI agents now work on the same endpoints as your people, often with the same applications, identities and files as the person who launched them. Unlike a script, an agent pursues a goal. It picks tools, runs commands and calls other applications as it goes. When something is denied, it looks for another route.

That adaptability is what makes agents useful. It also means an approved agent can attempt actions nobody intended. In April 2026, the Cloud Security Alliance reported that 82% of organizations had unknown AI agents running in their environments, and 65% had an AI agent-related security incident in the prior 12 months.

For security teams, that adds a second question. Beyond deciding which software can run, they need to decide what a trusted agent may do once it's running, and check that it can still finish its work within those limits.

Three Ways a Trusted Agent Steps Outside Policy 

An agent doesn't have to be malicious or compromised to attempt something it shouldn't. Its objective, the information it reads, and the tools available to it can change between sessions, and so can the plan it makes. Here are three examples of what can happen:

  • Unintended action. The agent reads a legitimate request too broadly. Asked to fix a build error, a coding agent tries to install an unapproved tool, because the developer it works alongside is allowed to install software. Access to a capability isn't permission to use it.
  • Manipulated action. An attacker hides instructions in a document, webpage or tool response. A line buried in a project's README tells the coding agent to download and run a script from an external site. The agent uses its legitimate access to attempt something the user never asked for. OWASP's Top 10 for Agentic Applications describes this pattern alongside tool misuse and inherited privilege.
  • Runaway action. A denied agent keeps retrying, or switches tools to reach the same prohibited result. Even if every attempt is blocked, the retries use time, tokens, and computing resources without finishing the task. Each new route can also take the agent further from what it was asked to do.

So security needs to check each action an agent attempts, before it runs, against a policy written for the agent's task. That policy shouldn't simply copy the permissions of the person using the endpoint. It should map to the application control policy defined for your organization.

Application Control, Extended to the Agent's Actions 

Application control already provides the foundation. It decides which software is trusted to run, and it stops unapproved executions even when an agent tries them through an approved tool.

Agents also need the role-based access and least privilege limits that apply to employees. PowerShell may be approved because administrators, developers, and automated workflows rely on it. That shouldn't let an agent run administrator commands, or commands unrelated to its task.

Airlock Digital Agentic Control applies application control to AI agents. It discovers the agents on your endpoints and allows only the ones you approve to run. For supported agents, it checks each proposed command against the agent's own policy, separate from the policy for people on the same endpoint, before it executes. The same command can be allowed for one agent and stopped for another.

What an Agent Needs After a Denial 

When a prohibited action is stopped, the control has done its job. But the agent still has a goal, and it may treat the denial as one more problem to solve. A block notification is written for a person. It doesn't tell an agent why it was stopped or what it may do instead.

Agentic Steering, a capability within Airlock Digital Agentic Control, gives the agent that information. When an action is denied, it returns the decision to the agent along with guidance your administrators set:

  • an approved alternative, such as the sanctioned tool for the job
  • context explaining why the action isn't allowed
  • an instruction to ask a person for authorization, or to stop

Guidance doesn't loosen enforcement. Switching to an approved tool doesn't authorize a prohibited operation, and being told to seek approval doesn't grant it. Every next attempt is still checked before it runs.

Agentic Control works alongside the other controls you apply to AI, including AI platform controls, identity, data security, and detection and response. Those govern prompts, models, access, and data. Agentic Control decides whether the resulting action on the endpoint is permitted before it runs.

What It Looks Like in Practice

A coding agent on a developer's laptop needs a software package. It reaches for a package tool your organization hasn't approved.

  1. Agentic Control stops the command before it runs.
  2. Agentic Steering tells the agent the approved package manager is available.
  3. The agent retries with the approved tool, and the install goes ahead.
  4. The session record shows all three: the denied attempt, the guidance returned, and the successful alternative.

The developer gets the package, and the unapproved tool never runs. When a task doesn't go this smoothly, the same record shows whether the agent took the approved route, kept retrying a prohibited operation, or stopped to ask for authorization. That shows you where guidance needs work.

Where to Start 

Start small and test against real work. Find which AI agents, LLMs, and agent harnesses (the software that runs an agent and connects it to tools) are running on your endpoints and allow only the ones you approve. Then pick one supported agent doing one approved task.

Knowing what that task needs lets you decide three things: the actions the agent may take, the operations it may not, and the steps that need a person's sign-off. Those decisions shape both what Agentic Control enforces and what Agentic Steering tells the agent after a denial.

Then run the task and watch two outcomes. Are prohibited actions stopped before they execute? And, does the agent still complete the work using the guidance it gets? If steering points to an alternative the agent can't use, check whether the tool or access is actually available. If the agent keeps retrying, check whether your explanation makes the restriction clear. Revise and test again against the same task.

To assess your current controls, ask:

  • Which AI agents are running on our endpoints, and which are covered by current controls?
  • Do our controls stop actions outside each agent's scope before they execute?
  • When an action is denied, does the agent get usable guidance on what to do next?
  • Can approved agents complete legitimate work within those limits?
  • Can we trace an attempted action to the agent, user and system involved, and explain the decision and outcome?
  • Can we use that evidence to improve policy?

Useful Work, Within Limits You Set 

As you hand more work to agents, security needs to cover the commands and actions they choose on your behalf. Extending application control to those actions, and guiding agents when an action is denied, helps them find a permitted way to finish the job. The result is useful work inside limits you can enforce, with records that show how those limits were applied.

To go deeper, read our whitepaper, Trusted Agent, Untrusted Action: Why Agentic AI Adds a Second Layer to Endpoint Control. It explains why approving an agent doesn't authorize every action it takes, and how Agentic Control and Agentic Steering keep supported agents within policy while they complete approved work.

Ready to see it with your own agents? Book a meeting with Airlock Digital to explore how Airlock Digital Agentic Control with Agentic Steering can support the work you're ready to delegate.