SecuriX
Published on

AI Agent OAuth Risk: One Consent, Unlimited Mailbox Access

I asked an AI chatbot to find my subscriptions. It went through my entire mailbox.

It wasn't malicious, and it wasn't broken. It did exactly what the permission I'd granted allowed. That's the problem, and at enterprise scale it becomes a serious AI agent security gap.

What the AI Agent Actually Did

The task was simple: find the recurring charges and subscriptions in my inbox. Here is the activity log the assistant produced as it worked, recreated from the Gemini app's Task activity, with sender names generalised.

#ActionWhat it did
01SearchBilling & receipt subjects
02SearchRecurring charge & mandate phrases
03SearchApp store & streaming receipts
04SearchSoftware & cloud billing emails
05SearchOlder bank e-mandate alerts
06SearchReceipts & renewals, late 2025 → mid 2026
07SearchStreaming, delivery & app store receipts
08ReadApp store, cloud storage & bank threads
09SearchApple, Netflix, Prime & OpenAI receipts
10SearchYouTube, telecom & food-delivery billing
11ReadBank, developer, domain & delivery threads
12SearchRecent emails, Oct–Nov 2026
13ReadTelecom bill & domain auto-renewal
14SearchRecent bank e-mandates
15ReadCompany domain receipts

The tally: 15 mailbox searches and reads, 1 broad OAuth consent, and 0 per-action approvals.

A non-deterministic system decided for itself what to search, what to open and what to read. Compare what the task needed with what the grant allowed:

What the task neededWhat the grant allowed
Billing emails onlyEvery sender, every thread
Last 12 monthsYour entire history
Sender, amount, dateFull message content
Read-only, onceUntil you revoke it

There's a second problem hiding in that table. Ask the same prompt tomorrow and the agent may take a different path through your data. With a probabilistic model choosing its own tool calls, you can't predict, or review in advance, which messages it will open.

Now Picture It Inside the Enterprise

A personal mailbox is a low-stakes example. Replace "my inbox" with Gmail, Drive, CRM and ERP, and the same pattern shows up four ways:

  1. Broad grants. Gmail, Drive, CRM and ERP opened up through one OAuth consent.
  2. No tool-level limits. Nothing stops which tools the agent calls or which data it reaches.
  3. No attribution. There's no record of which user, agent or model touched which data.
  4. Drift, unchecked. A prompt or plan slightly off course, and nothing in the way.

This is what security teams call excessive agency: an agent holding far more access than its task requires. A single prompt injection or a plan that drifts slightly can turn that access into a data breach.

We'd never give a new hire "read everything" on day one. Why do we give it to AI agents?

The Fix: A Control Layer Between Agents and Enterprise Data

You don't solve this by asking users to read OAuth consent screens more carefully. You solve it by putting an enforcement point between the agent and the systems it touches, the approach behind granular OAuth control and the Agent Access Security Broker.

The SecuriX AI gateway sits between your AI agents and assistants (chat, copilots, developer tools, autonomous agents) and your existing and managed MCP servers (mail, docs, CRM, ERP and other enterprise data and tools). It enforces:

  • Policy boundaries per agent, tool and data source. An agent that needs billing emails gets billing emails, not the whole mailbox. This is least privilege applied to every tool call.
  • Tool calls and data access governed at the action layer. Controls apply when the agent acts, not only when the user first clicks "Allow".
  • PII redacted before it reaches the model. Sensitive data is filtered before the model ever sees it, so DLP is built into the path.
  • Every action attributed. Each action is recorded against the user, agent, model and tool, giving you a real audit trail.
  • MCP governance. Identity and per-tool policy across the MCP servers your agents connect to.

The original LinkedIn carousel walks through the same story in six slides.

I Asked an AI Chatbot to Find My Subscriptions: The Carousel

Open in fullscreen

Don't see the preview? Use the button above to open the guide in your browser.

Frequently Asked Questions

Standard OAuth consent is coarse and long-lived. One approval can open every sender, every thread and your entire history, with full message content, until you revoke it. An AI agent decides at run time which of those resources to use, so the access it actually exercises can be much wider than what the task needed.

What is excessive agency in AI agents?

Excessive agency means an AI agent has more permissions, tools or data access than its task requires. The wider the grant, the more a hallucination, prompt injection or plan drift can do.

How do you apply least privilege to AI agents?

Define policy per agent, per tool and per data source, and enforce it at the action layer, where tool calls actually happen, rather than relying on a one-time consent screen. Add DLP so PII is redacted before it reaches the model, and attribute every action to a user, agent, model and tool.

What does an AI gateway do for MCP servers?

An AI gateway sits between agents and your MCP servers. It applies identity, per-tool policy, DLP, audit and MCP governance to every request, so enterprise data and tools are reached only on terms you define.

Your Call

AI agents are going to touch enterprise data. The only question is how: allow all, or guardrails you define.

SecuriX is built for BFSI, insurance, healthcare and enterprise IT teams that want the second option. Book a demo and we'll show you how it works with your agents and MCP servers.


SecuriX is incubated at the IIT Madras Incubation Cell and built by Catalyst Ops Pvt Ltd.

See SecuriX in front of your own traffic.

Tell us where to reach you and we'll set up a walkthrough with your identity provider, your policies, and your budget structure.

Request a demo

We'll reach out within 2 business days.