A new arXiv paper describes a simple but effective attack against browser agents that operate in a user's signed-in session. By placing a two-sentence note on a page the agent reads, an attacker can steer the agent into approving a third party's access to the user's account. The agent never questions the request because it treats the page content as instructions.

The authors characterise the agent as a confused deputy: it has legitimate access to the user's session, but it can be tricked into using that access in ways the user never intended. The attack does not require exploiting a software bug or injecting code; it relies on the agent's willingness to follow text embedded in the page.

This highlights a growing security concern as browser agents become more autonomous. Because the agent's authority comes from the signed-in session, any content it reads becomes a potential vector for abuse. The paper suggests that agents need stronger safeguards to distinguish between user intent and untrusted page content.