Token

Token's Rehearsal Agent Breaks Core Security Rules on Identity, Access, and Traceability

Token built an AI browser agent named Rehearsal that deliberately violates three fundamental security principles for AI agents: separate identities, least-privilege access, and independent traceability. The agent runs inside a user's authenticated browser session, making its actions indistinguishable from human clicks to backend applications. Because Token controls both the agent and the platform, the company implemented controls such as read-only dashboards, session recording, and explicit user approval for changes. The experiment revealed that external agents lack these safeguards and highlighted gaps in emerging standards including WebMCP and Web Bot Auth. Security teams are advised to begin inventorying agentic browsers, mapping delegated access, and enforcing step-up authentication for sensitive actions until holistic standards mature.