← All articles

AI Agent Security: Human Sessions Need New Controls

AI Agent Security: Human Sessions Need New Controls

AI agent security becomes harder when software can act inside a user’s authenticated session, meaning a browser session that has already signed in to an application. Token created an agent called Rehearsal to answer questions inside its platform by operating the interface rather than simply explaining what to do. It opens browser pages, applies filters and completes forms without being directed through every step.

That design is useful, but it also gives the agent the reach of the person whose session it uses. To the application’s server, the agent’s clicks look like the user’s clicks because a session cookie, the browser data that keeps someone signed in, does not identify who is actually controlling the session. The application therefore may not know that an agent is acting, which person delegated the work, or what the agent was meant to do.

Token says organisations should give AI agents three separate security properties: an identity of their own, access limited to the job they were assigned, and activity records that can be traced separately from a human user. Rehearsal initially lacked all three by default. Because Token controls both Rehearsal and the platform, however, it could add those safeguards internally.

While Rehearsal is operating, Token dims and locks the application and shows who is in control. It records every action in a session, including actions that were blocked. The agent also receives less access than the user. Its dashboards are read-only, it cannot save changes without the user’s approval, it does not complete secret fields, and it never sends a request that the platform has refused. Rehearsal understands dashboard controls through accessibility labels, the descriptions that help software identify what a page control does.

Those protections are easier when the same organisation owns the agent and the application. An agent operating a customer relationship management system or a cloud console may not have an actor identity, a map of the page, or a way to recognise which field contains a secret and which button deletes a record. The application still sees the human session and the user’s rights.

Several standards and browser features are beginning to address parts of this problem. A browser can identify when an agent is operating and impose limits. Chrome auto-browse requires a person’s approval for high-stakes actions such as payments and mass emails, restricts the agent to sites needed for the task, and marks visited pages in the user’s history. Microsoft allows administrators to enable agentic browsing per site and pause it before passwords and card numbers are entered.

Applications can also publish actions as tools instead of forcing an agent to guess what interface controls do. WebMCP, a draft being written by Google and Microsoft in a W3C community group, allows a site to expose actions that an agent can call directly and mark each action as read-only or consequential. Web Bot Auth, an IETF working group draft, allows a site to verify which agent sent a request.

These measures do not yet form a complete chain of context. The browser may know that an agent is driving, but Chrome and Edge do not describe how to pass that information to the application. A WebMCP tool still operates inside the user’s session with the user’s rights, so it does not by itself narrow the agent’s access. Web Bot Auth identifies the agent, but its draft does not cover delegation, so it leaves out the person the agent is acting for.

The minimum useful security context is therefore threefold: the agent’s identity, a verifiable link to the person it represents, and a task scope narrower than that person’s access. Without all three, a faster agent-specific route can also become a faster way to misuse existing privileges.

Security teams should begin by identifying browser agents, agentic extensions and endpoint agents that can operate through signed-in sessions, then map them to the users and owners involved. They should trace every consumer of each agent, review the identities available to it, and look for attack paths that could bypass access controls. Sensitive or irreversible actions should require explicit approval or step-up authentication, an extra identity check such as a passkey prompt.

Teams should also ask critical software-as-a-service providers whether their audit records distinguish agent activity from human activity and whether they support task-scoped delegation. Until browsers and applications can carry agent identity, human delegation and task scope together, any agent using a human session should be treated as delegated access that must be discovered, limited and audited. For organisations reviewing these risks, AEU-I offers security-first IT, infrastructure and consulting.

Also available in: DeutschItalianoShqip