Safety & Security

Airlock Digital adds action-level controls for Codex and Claude

Airlock Digital's version 7.1 evaluates proposed Codex and Claude actions before execution, blocks disallowed commands and returns explanations or permitted alternatives for policy-aware retries.

Airlock Digital Launches Agentic Steering to Control AI Agent Actions Before Execution
Airlock Digital Launches Agentic Steering to Control AI Agent Actions Before ExecutionInnovalabs / pixabay
By James Calloway6 min read

Why it matters

  • Airlock Digital launched Agentic Steering in version 7.1 for Codex and Claude.
  • Denied actions can return an explanation or an approved alternative for a policy-aware retry.
  • Version 7.0 introduced discovery for AI agents, language-model tooling and harnesses across endpoints.
  • Version 7.1 records sessions, commands, files, policy decisions, outcomes, risk activity, token usage and estimated cost.
  • The capability was previewed at Black Hat USA 2026 as Agentic AI Control & Governance.

By James Calloway

Airlock Digital introduced Agentic Steering in version 7.1 of its endpoint security platform for Codex and Claude. Agentic Steering evaluates an AI agent’s proposed action before execution, permits actions that satisfy organizational rules, and returns an explanation or an approved alternative when it denies one. An AI agent is software that can pursue a task by invoking applications and changing its approach; endpoint security software governs what runs and does on employee devices.

The launch addresses an enterprise control gap. An approved assistant can operate other software, work with files, and pursue a task in ways its buyer did not directly inspect. Approval of the application therefore does not decide whether every command should run.

What does agentic steering change?

Traditional application control (a policy for deciding which software may run) answers whether software may run on a device. AI agents add another layer: once started, they can invoke applications, work with files, and revise their approach. On a shared workstation, a person and an agent may otherwise inherit access to many of the same resources.

Airlock Digital’s Agentic Control product page describes separate, centrally managed policies for people and agents on the same endpoint. Agentic Steering evaluates proposed actions from supported agents against organizational rules, permits authorized actions, and stops prohibited ones.

That is more specific than asking whether the agent itself appears on an approved-software list. A company may authorize an assistant for development work while restricting particular commands or administrative operations it attempts. The rules must follow the actor and action, rather than treating every operation from a trusted application as equally acceptable.

“Traditional application control determines what software is trusted to execute. Agentic AI introduces a new actor on the endpoint, with an ability to execute other applications,” said David Cottingham, Airlock Digital’s co-founder and chief product officer. “Teams now need the ability to define and enforce what an AI agent is allowed to do once it’s running.”

Takeaway: In plain terms, Airlock is separating permission to launch an agent from permission for each action it takes.

How does a denial become useful to an agent?

A conventional block notification may tell a person what went wrong without giving an autonomous tool enough information to choose a safe next step. An agent can interpret failure as a reason to try a different route toward its original objective.

Agentic Steering returns the decision directly to the supported agent. Administrators can attach context explaining why an action was denied or identify an alternative that the organization permits. The intended result is a policy-aware retry, rather than a sequence of increasingly improvised attempts.

For example, an organization could point an agent toward an approved tool when a requested command is blocked. The enforcement decision remains outside the agent: the explanation helps it plan, but it does not grant permission to bypass the rule.

That boundary matters. Guidance only helps if the agent interprets it correctly and an approved alternative can complete the task safely; the announcement offers no measured retry rate, evidence of fewer policy violations, or detail on conflicting rules.

Takeaway: A denial becomes actionable only when the response explains both the constraint and a permitted route forward.

What can administrators see after an agent acts?

The release builds on AI discovery capabilities introduced in Airlock Digital 7.0. Security teams can identify AI agents, language-model tooling, and harnesses across endpoints, then use deny-by-default application control to restrict execution to approved software. A harness is the layer that connects a language model to tools, tasks, or an execution environment.

Agentic Steering adds action-level evaluation for the supported agents. Together, the features separate two questions: whether an AI application should run at all, and whether its proposed behavior is permitted once it starts.

Airlock Digital also centrally records:

  • Agent sessions
  • Commands
  • Files
  • Policy decisions
  • Outcomes
  • Risk activity
  • Token usage, the volume of model input and output
  • Estimated cost

Teams can search those sessions and review AI-generated summaries of what an agent attempted and how enforcement affected the result. Frequent attempts to use an unapproved tool might reveal an unclear workflow or a missing approved alternative. Reviewing later actions can show whether the agent used the guidance supplied after a denial.

The records can inform operational policy changes. They do not establish that the product retrains an underlying AI model, and the announcement does not specify how it validates the generated summaries.

Takeaway: The audit trail can improve rules, but it records agent behavior rather than changing the model.

How does this fit the existing platform?

Airlock’s application allowlisting technology (controls that permit only approved applications, scripts, and processes) already governs execution. Published trust rules work at the file, path, publisher, and parent-process levels, with exception workflows in place. A parent process is the program that starts another program.

Agentic Steering applies that preventative approach to a new source of endpoint activity. The wider platform supports Windows, macOS, and Linux and integrates with existing security and IT workflows. Its endpoint coverage is broader than the launch’s stated agent coverage.

The launch specifically identifies Codex and Claude. It does not promise identical action-level control over every AI tool, and it does not provide a schedule for additional supported agents.

Airlock positions endpoint execution control as complementary to identity security, data protection, AI platform controls, and detection and response. Checking whether a command may execute does not establish that a model response is accurate or that every use of data is appropriate. Those systems answer different questions about an agent’s operation.

Takeaway: Action control complements other security layers, but it cannot certify model accuracy or appropriate data use.

When is it available, and what must buyers prove?

Agentic Steering is available to Airlock Digital customers with version 7.1 for Codex and Claude. The company previewed the capability at Black Hat USA 2026 under the name Agentic AI Control & Governance.

For enterprises adopting autonomous tools, approving an agent marks the beginning of access control rather than its conclusion. Airlock is trying to connect execution policy with guidance the agent can use, placing security boundaries inside the loop between an objective, a proposed action, and its result.

The release names no customer deployments, measured reductions in risky actions, or comparisons with competing controls. The next meaningful proof point will be operational evidence that denied agents choose supplied alternatives and remain within policy, rather than simply generating more blocked paths.

Takeaway: Airlock has released a defined control for Codex and Claude, while its enterprise value still depends on proven behavior in real deployments.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. Cato Networks Launches AgentSense to Block AI Agent Attacks
  2. 83% of Firms Run Unapproved AI Coding Agents, BlueFlag Research Finds
  3. AI Agents Move Onto Windows PCs, and Permissions Become the New Battleground
  4. OpenAI details how it runs Codex safely in production
  5. Tanium launches AgentSight to track and govern AI tools on endpoints

« Previous articleNext article »