From Xenovia LabsIssue 3, October 2026
Secure Enough

Probably fine.

ReportMachines

The permission problem

AI agents are being handed the keys to email, code, calendars, and money. Almost nobody is asking which keys, or why.

By Desmond Faltz, Prompt Injection Desk4 min read

The pitch for an AI agent is that it can act. Not answer, act: read your inbox and reply, open a pull request, book the travel, file the expense, update the ticket. Every one of those verbs requires access to a system, and over the past year that access has been granted at a pace and breadth that would get a new human employee a meeting with the security team.

We spent six weeks talking to platform engineers, identity administrators, and the people who build agent integrations. The picture that emerged is less a scandal than a shrug: agents are being connected to production systems with whatever permissions were easiest to obtain, by people who are not thinking of the agent as a user, through connectors that often cannot express anything finer than "everything."

How agents get their keys

An agent acts through tools: connectors to email, calendars, document stores, code hosts, databases, ticketing systems, payment platforms. In practice, those connectors are authorized in one of three ways.

  • Delegated user credentials. The agent acts as you, with your OAuth token or session. Whatever you can do, it can do, including the things you would never do.
  • Service accounts. Someone creates a dedicated identity for the agent. In the deployments we reviewed, these were routinely granted admin roles "to avoid permission errors during setup" and never trimmed afterward.
  • Long-lived API keys. Pasted into a configuration file, scoped to the whole account, rotated never.

The connectors compound the problem. Many tool protocols expose a server's full capability set to the model at once; if the mail connector supports read, send, delete, and forward, the agent has all four, regardless of whether the task is "summarize my morning."

Why this is different from a user with too much access

A human with excess privilege is a risk. An agent with excess privilege is a risk multiplied by everything it reads.

Agents consume untrusted content constantly: emails from strangers, web pages, documents, ticket descriptions, calendar invites. Any of that content can contain instructions, and current models cannot reliably distinguish instructions they should follow from text they should merely process. This is prompt injection, and a year of mitigation attempts has established that it is a property of the architecture, not a bug awaiting a patch.

Security researchers have converged on a simple framing for when this becomes a breach, often called the lethal trifecta. An agent is dangerous when it simultaneously has:

  • access to private data,
  • exposure to untrusted content, and
  • a way to send data somewhere else.

Most agent deployments we looked at had all three on day one. The mail agent that can read everything, receives mail from anyone, and can send mail to anyone is the canonical example, and it is also the most common product on the market.

"We would never give a contractor a token that can read every mailbox and send as anyone. We gave it to a chatbot because the setup guide told us to." — Identity engineer at a mid-sized software company

The confused deputy returns

There is an older name for part of this problem: the confused deputy. A program with legitimate authority is tricked into using it on behalf of someone who does not have it. Agents are confused deputies by construction. They hold your authority and take direction, in part, from whoever can get text in front of them.

The reframing matters because it points at the fix. You cannot make the deputy un-confusable with better instructions. You can make it hold less authority.

What good looks like

The organizations that had thought this through shared a few habits.

  • Treat the agent as a user, with its own identity. Not your token. A named identity that appears in audit logs, can be disabled in one place, and has permissions reviewed like anyone else's.
  • Scope by task, not by product. A triage agent gets read-only mail. A drafting agent gets create-draft, not send. Sending is a separate, human-approved step.
  • Separate read from write from send. Where the connector cannot express that, put a policy layer in front of it, or do not connect it.
  • Human approval for irreversible and outbound actions. Payments, deletions, external messages, permission changes, merges to protected branches.
  • Short-lived credentials. Tokens that expire in minutes limit the damage of any single injection.
  • Logs the agent cannot touch. Every tool call, every argument, somewhere the agent has no write access.
  • Break the trifecta deliberately. If an agent must read untrusted content and private data, remove its outbound channels. If it must send, do not let it read from strangers.

None of this is novel security thinking. It is least privilege, applied to a new kind of principal that happens to be extremely persuadable. The novelty is only that we are currently skipping it.

About the writer

Desmond Faltz

Prompt Injection Desk

Desmond covers language models, agents, and the growing number of ways to make a very expensive computer do something it was asked not to. He has never once ignored his previous instructions.