Airlock
An enterprise-oriented AI agent governance product originating from internal experiments at Atlas. It provides IT and security teams with a unified framework for governing agent behavior, supporting real-time control, policy auditing, and human-confirmation mechanisms.
Timeline (last 90 days)
At the init() 2026 conference, WorkOS showcased three products—Airlock, Atlas, and Passport—in the form of live previews.
Airlock, Atlas, and Passport are currently in the preview stage and still some distance from production availability, requiring waiting for open trials or official release.
The simultaneous launch of Airlock, Atlas, and Passport shows that WorkOS is moving toward a more complete enterprise infrastructure platform, not content to be just an SSO vendor.
WorkOS's official recap did not provide detailed feature lists, APIs, pricing, or technical architecture information for the three products Airlock, Atlas, and Passport.
Airlock inserts policy as a checkpoint between the agent and actual execution, requiring every call to first pass an intent-level review, shifting the security boundary from post-hoc auditing to pre-emptive interception.
WorkOS used an email-sending demo to illustrate how policy translates into actual action: an agent's request to send an email is first evaluated by Airlock against preset policies, and the call is only executed when the intent matches the policy; otherwise it is blocked.
The Airlock system includes a component called Pipes, which is responsible for carrying the actual routing channel of agent calls, handling request routing and governance.
WorkOS launched Airlock, introducing the concept of 'intent-based access control,' which evaluates whether the intent behind a call made by an agent conforms to preset policies at the time the agent initiates the call.
Airlock adopts a three-layer governance structure: deterministic rules, AI judgment, and human approval.
When an action cannot be directly permitted or denied by a hard rule, Airlock can invoke AI to assess its risk and intent.
14 more timeline events
All Facts (20)
Airlock provides a solution for adding human approval steps to AI Agent tool calls, centered around three core capabilities: reviewing the exact request, resuming execution, and handling edits and retries.
60%UnverifiedAirlock embeds governance into the Agent execution pipeline for real-time control, and can decide to allow, block, or require human confirmation based on preset rules.
60%UnverifiedWorkOS's official recap did not provide detailed feature lists, APIs, pricing, or technical architecture information for the three products Airlock, Atlas, and Passport.
50%UnverifiedThe simultaneous launch of Airlock, Atlas, and Passport shows that WorkOS is moving toward a more complete enterprise infrastructure platform, not content to be just an SSO vendor.
50%UnverifiedAirlock inserts policy as a checkpoint between the agent and actual execution, requiring every call to first pass an intent-level review, shifting the security boundary from post-hoc auditing to pre-emptive interception.
50%UnverifiedThe Airlock system includes a component called Pipes, which is responsible for carrying the actual routing channel of agent calls, handling request routing and governance.
50%UnverifiedWorkOS used an email-sending demo to illustrate how policy translates into actual action: an agent's request to send an email is first evaluated by Airlock against preset policies, and the call is only executed when the intent matches the policy; otherwise it is blocked.
50%UnverifiedAirlock adopts a three-layer governance structure: deterministic rules, AI judgment, and human approval.
50%UnverifiedWhen an action cannot be directly permitted or denied by a hard rule, Airlock can invoke AI to assess its risk and intent.
50%UnverifiedFor the most sensitive operations, Airlock retains a human approval step.
50%UnverifiedThe name Airlock borrows from the concept of an airlock chamber in aerospace engineering, referring to the establishment of a controllable buffer checkpoint between the Agent and the real system.
50%UnverifiedAirlock allows approvers to directly modify request content before approving, without needing to send it back for the Agent to regenerate.
50%UnverifiedAirlock allows approvers to see the exact content of the request the Agent is about to execute, including the specific tool being called, the parameters passed in, the target endpoint, and other complete information.
50%UnverifiedAirlock supports seamlessly resuming Agent execution from the paused point after human confirmation, preserving context state rather than interrupting and restarting.
50%UnverifiedAirlock is an MCP authorization governance tool that can enforce tool permissions, inspect request content, and test for data leakage.
50%UnverifiedAirlock is a control layer for governing merges that inserts a controllable isolation barrier between approval and the actual merge.
50%UnverifiedAirlock's most important rule is to prevent agents from merging code that has changed after approval.
50%UnverifiedAirlock is a product that provides IT and security teams with a unified governance framework for Agent behavior.
50%UnverifiedAirlock originated from an internal experimental project called Atlas.
50%UnverifiedAt the init() 2026 conference, WorkOS showcased three products—Airlock, Atlas, and Passport—in the form of live previews.
50%