The Hacker News
Token Tests How Its AI Browser Agent Bypasses Security Rules
Token reveals how an AI browser agent, dubbed Rehearsal, was built to deliberately bypass three core security rules—identities, job-specific access, and task-scoped traceability—and what that test proves for wider deployments. In their write‑up, Token explains that Rehearsal runs inside the user’s browser session, performing actions by clicking, filtering, and filling forms rather than merely advising.
To the server, the agent’s clicks look indistinguishable from a human’s, because the session cookie does not disclose the driver’s identity. Token notes that, because the app owns both the agent and the platform, internal controls can be configured to reveal who is driving the session and to limit the agent’s actions, but outside agents would lack that visibility.
The team suggests practical measures: dimming or locking the app while the agent operates, recording every move (including blocked ones) per session, and restricting the agent’s reach to read-only dashboards or to actions that require explicit user approval, with safeguards such as accessible‑label guidance for widgets.
The piece argues that the real challenge is cross‑layer context: browsers, agent frameworks, and apps each emit different signals about identity, delegation, and scope, yet none fully conveys the whole picture to the other side.
It cites ongoing or proposed developments—Chrome’s agent‑based browsing, Microsoft’s per‑site agent controls, Google and Microsoft’s WebMCP drafts for exposing agent actions with read-only vs. consequential flags, and the Web Bot Auth effort for verifying agent provenance—as essential pieces that lack a unifying, interoperable layer.
The authors recommend security teams begin inventorying agentic browsers and extensions, map agent activity to users, enforce step‑up authentication for sensitive actions, and demand that SaaS vendors confirm audit records can distinguish agent from human activity and support task‑scoped delegation. The note closes by urging organisations to anticipate standards while tightening internal controls, given that agent‑driven sessions will proliferate before formal guidance arrives.