SIAgentNews.com
← All news

Airlock Digital launches Agentic Steering to govern AI agent actions

The new security tool allows organizations to evaluate and guide specific commands made by AI agents like Claude and Codex before they are executed.

· First reported by Unite.AI

7.1Airlock Digital platform version

The short version

  • The tool evaluates specific actions from AI agents rather than just trusting the application as a whole.
  • It currently supports actions from Codex and Claude within version 7.1 of the platform.
  • Administrators can provide agents with approved alternatives or explanations when a command is blocked.
  • The system records sessions, including token usage and costs, to help teams refine security policies.

Why it matters

As companies move from using AI as a chatbot to using it as an autonomous agent that can run software, they need a way to set boundaries. This technology ensures that an authorized AI tool cannot perform unauthorized administrative tasks or access restricted files.

Moving beyond app trust

Traditional security models usually focus on whether a piece of software is allowed to run on a computer. If a company trusts an application, that software is generally given the permissions it needs to function. However, AI agents introduce a new layer of complexity because they can perform many different tasks once they are active.

An AI agent can invoke various other applications, work with different files, and change its strategy as it tries to reach a goal. This means that simply trusting the AI application does not mean every single command it generates should be permitted. Airlock Digital is addressing this gap by allowing administrators to define specific boundaries for what an agent is allowed to do.

The new Agentic Steering feature allows organizations to distinguish between the agent itself and the actions it takes. A company might authorize an AI assistant for development work but still want to block it from performing specific administrative operations. This shift moves security from a broad 'yes' for an app to a granular 'yes' for specific behaviors.

How Agentic Steering works

When a supported AI agent proposes an action, Agentic Steering evaluates that specific request against the organization's established rules. If the action is permitted, it proceeds as planned. If the action is prohibited, the system stops it and sends a decision back to the agent.

Unlike a standard security block that might just show an error message to a human, this system provides context to the AI. Administrators can attach an explanation for the denial or suggest an approved alternative that the agent is allowed to use. This allows the agent to understand why it failed and attempt to complete its goal using a different path.

The goal is to create a policy-aware retry loop rather than letting the agent guess what to do next. By providing guidance, the system helps the agent stay within the rules without the agent having to improvise. The enforcement remains outside of the agent, ensuring it cannot bypass the rules even if it tries to find a workaround.

Visibility and session review

The new capability builds on previous discovery tools that help security teams identify AI agents and language-model harnesses across their endpoints. Once these tools are identified, teams can use a deny-by-default approach to ensure only approved software can run. Agentic Steering adds the next step by evaluating the behavior of those approved tools.

Airlock Digital centrally records the entire session of an AI agent's activity. This includes the commands the agent tried to run, the files it accessed, and the final outcomes of its actions. It also tracks risk activity, token usage, and the estimated cost of the agent's operations.

Security teams can review these AI-generated summaries to see how enforcement affected the agent's progress. If an agent frequently tries to use an unapproved tool, it might signal that a workflow is unclear or that an approved alternative needs to be created. These records help administrators refine their rules based on real-world usage data.

Platform integration and scope

Agentic Steering is part of the wider Airlock Digital endpoint security platform, which supports Windows, macOS, and Linux. The platform already includes features for controlling scripts, processes, and files based on publishers or parent processes. This new tool extends those preventative security methods to the specific activity of AI agents.

While the platform is broad, the current launch specifically identifies support for Codex and Claude. The company does not currently promise identical control over every AI tool on the market. This distinction is important for organizations looking to understand which specific agents they can govern with this technology today.

Airlock Digital positions this tool as one part of a larger security strategy. It is intended to complement identity security, data protection, and detection and response systems. While it checks if a command can execute, it does not guarantee that the AI's response is accurate or that the data it uses is appropriate for the task.

The future of agent governance

The launch of Agentic Steering makes a practical distinction for companies adopting autonomous tools. It establishes that approving an AI agent is only the first step in access control, not the final one. By connecting execution policy with guidance, the system makes security boundaries a part of the agent's working loop.

As agents become more common in the workplace, the ability to manage their behavior will become a standard requirement for IT departments. Organizations need to ensure that these tools can pursue objectives without creating unintended risks or violating internal policies. This technology provides a framework for that balance between autonomy and oversight.

The capability was previously previewed at Black Hat USA 2026 under a different name. Now that it is officially launched in version 7.1, it provides a way for teams to define what an AI agent is allowed to do once it is running. This ensures that the agent remains a productive tool rather than an unpredictable actor on the corporate network.

What to watch

  • Which other AI agents will receive support beyond Codex and Claude?
  • How will organizations balance the need for agent autonomy with the need for strict security boundaries?

Try it

Check your organization's current AI policy to see if it defines specific actions agents can take, or if it only covers which AI apps are allowed to be installed.

Sources and links

First reported by Unite.AI. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news