We Built An Agent That Broke All of Our Security Rules. Here's What We Learned.

Researchers at Token built an AI agent that intentionally violated their own security guidelines, which normally recommend unique identities, limited access, and separate traceability for such agents. The agent was designed to break these rules by default, allowing the team to explore how to defend against similar agents. This experiment helped the researchers think through defense strategies for agents that do not follow standard security protocols. The agent in question was a browser agent that operated within a user's authenticated session. The exercise provided valuable insights into securing AI agents that may not adhere to typical security rules. The findings are based on the team's experience with their own agent at Token.