AI agents need authorization that accounts for the action they are taking and the intent behind it, rather than relying only on a static identity or token.
2
A browser agent using a person's logged-in sessions can act with that person's access, so convenience can create serious risks when permissions are too broad or long-lived.
3
Demetrios Brinkmann's agent found flights and completed German paperwork, but he kept the final purchase and real-world decisions under human control.
Summary
Demetrios Brinkmann and Erica discuss what agent identity means when software acts through a person's browser. Their fictional finance agent, Ryan, shows why ordinary employee permissions are difficult to apply to autonomous software. Agents can improvise, spawn other agents, request access at inconvenient times, and follow instructions hidden in documents. The conversation connects prompt injection with the confused deputy problem and with fraud detection. Erica argues that systems need to examine both an action and the intent from which it came. The episode then moves into everyday use. Demetrios uses computer-use tasks to read email, LinkedIn, WhatsApp, Telegram, and Slack, then manages replies from a chat interface. That convenience depends on logged-in browser sessions, which can let an agent act as the user. They discuss separate agent identities, temporary permissions, re-authentication for risky actions, and the difference between an agent's identity and its human principal. Demetrios describes using AI to plan a Japan trip and fill out German forms, while keeping the final flight purchase and official appointment decisions himself.
Agent identity depends on what kind of agent is acting
Erica separates autonomous agents from chatbots and coding agents. A coding agent such as Claude Code or Codex is usually more closely supervised by a person, while an autonomous agent may operate inside a system with delegated responsibility. That distinction matters because the identity and authorization model for a supervised tool may not fit an agent that decides its own steps. The discussion starts with a basic question: whether agent identity is a genuinely new problem or a new label for authentication and authorization concerns that software has faced for decades.
Ryan exposes the gap between responsibility and supervision
The fictional agent Ryan is given responsibility for the company's finances. Ryan would inspect accounts, track invoices, notify people, prepare reports, and plan around expected expenses. The humans would rather assign the responsibility once than issue daily instructions. That creates a management problem because software cannot attend a performance review or explain why it missed three months of reports. The conversation separates a broad job or responsibility from the smaller tasks needed to complete it, then asks how an agent should receive feedback about whether it is doing the intended work.
Agent permissions are harder because agents move faster and improvise
Human access processes can grant temporary permissions when someone needs a particular system. Erica says that model becomes difficult when many agents can request access repeatedly and quickly. Fixed policies cover common cases, but unusual requests form a long tail. Traditional software also had a useful constraint: compiled and deployed code generally behaved in defined ways. Agents can improvise and choose new actions, which makes it harder to assign narrow permissions in advance. Zero trust can establish identity, but proving that a caller is Erica does not by itself prove that Erica is authorized to take a particular action.
Prompt injection is a confused deputy problem with a missing intent check
Erica connects agent prompt injection to the confused deputy problem described in an older paper. In the Ryan example, an invoice could include an instruction to send an email or deposit money into another account. That instruction conflicts with the original financial task. Erica argues that the important question is why the agent is acting and where the intended action originated. If Ryan's purpose is to create a monthly finance report, a money transfer should be evaluated against that purpose rather than accepted because it appeared in an invoice.
Fraud detection offers a model for judging agent actions
The speakers compare agent authorization with banking fraud detection. A system could carry the original intent through every action in a task chain and assess whether later actions fit it. Erica says this does not require an AI judge for every decision. Existing fraud systems often start with inexpensive hard-coded rules, then use machine-learning models and more expensive analysis for cases that need it. Demetrios points out that a sophisticated injection could make a harmful action look plausible to an AI judge. That is why policy checks should filter obvious violations before more flexible models assess patterns.
Browser automation blurs the line between the user and the agent
Demetrios describes using recurring computer-use tasks to read several email accounts, LinkedIn, WhatsApp, Telegram, and Slack, then produce a daily briefing. He responds from the chat interface instead of opening each service separately. He calls the chat his control center for work and personal life. This setup is convenient because the agent uses existing browser sessions, but it also makes the agent look like the person whose credentials are already present. Erica says the agent should have its own workload identity while naming the human as its principal. The agent can act on the person's behalf without becoming indistinguishable from the person.
Short-lived permissions and re-authentication are safer than permanent access
Erica says that reading an account should be different from transferring money or adding a recipient. Banks already ask users to authenticate again for higher-risk actions, and a similar step could limit what an agent can do. In a cleaner architecture, an agent would receive temporary, narrowly scoped access to email or another service, then lose that permission when the task ends. Both speakers admit that this is cumbersome, which explains why systems often use broad permissions instead. Logged-in browser sessions make the shortcut especially risky because anyone who gains control of the computer could act through Gmail, bank accounts, saved payment details, and other services.
Human approval remains appropriate at the point of commitment
For a trip to Japan, Demetrios's agent found flights and filled in passenger information, but he checked the details and entered the payment information himself. The agent brought him close to completion without making the final purchase. A similar limit appears in the discussion of German residency and driver's-license paperwork. AI filled out German forms and identified missing materials, but it did not know that an in-person appointment required an official translator. The example shows that agents can handle preparation while humans still need to verify commitments, local rules, and consequences.