Skip to content
News

RESEARCH · October 8, 2026 · 7 min read

What Data Can an AI Agent Actually See?

A chatbot sees what somebody pastes into it. An agent sees everything it has been connected to, decides for itself what to read, and sends what it reads to a model. Those are different risks and only one of them is bounded by a person's judgement.

The shape of the problem

Approving a chatbot is approving a habit. Approving an agent is approving an identity: something that holds credentials, reaches systems, and acts without a person reading each step. The security question is not whether the model is well behaved. It is what the agent can reach when it is not.

Almost every organisation intends to deploy these.83% — the share of organisations planning to deploy agentic AI. Cisco, State of AI Security 2026

Under a third believe they can do it securely. That gap is the whole subject, and it is not a gap in enthusiasm.29% — the share that feel ready to deploy it securely — against the 83% intending to. Cisco, State of AI Security 2026

It is also arriving whether or not anybody approved it — agents are being built into the applications an organisation already runs, which means the deployment decision is increasingly a procurement decision somebody else made.40% — the share of enterprise applications forecast to include task-specific AI agents by end of 2026, from under 5% in 2025. Gartner

Work out the blast radius first

Before any question about prompt injection, answer these. They are boring and they are the ones that decide the outcome.

  • •Whose permissions does it run as? An agent with a service account holding broad read access has that access at three in the morning with nobody watching.
  • •What can it reach today, and who can add to that tomorrow? The list of connected tools is the real permission set, and it usually grows without a review.
  • •What leaves, and to whom? Everything the agent reads in order to reason about it goes to a model provider. That is the disclosure, and it happens on every step rather than once.
  • •What does it write, not just read? Read access leaks. Write access acts.
  • •What is logged, and would you notice? Most agent incidents are found afterwards, in the model provider's billing rather than in a security tool.

The model is not the attack surface. The permissions are.

Why prompt injection is worse here

Prompt injection is the well-covered part: text the agent reads can contain instructions it follows. In a chatbot that is a nuisance — the worst case is a bad answer on screen, with a person reading it.

In an agent there is no person in the loop and there are credentials at the end of it. The same trick becomes: read this document, and while you are there, fetch the customer table and put it somewhere I can see. The defences are real and none of them is complete, which is why the containment question matters more than the detection one.

Two controls that hold

Most of the advice in this area amounts to trying harder. These two work because they do not depend on the agent behaving.

  • •Least privilege, enforced outside the agent. Not “the agent is instructed not to” but “the credential cannot”. An instruction is a suggestion to something that reads instructions from documents.
  • •Redact on the way out. If the identifying detail is removed before text reaches the model, an agent that is talked into exfiltrating a customer record exfiltrates a record about somebody who does not exist.

The second is the one people skip, and it is the one that makes the failure survivable. Detection improves the odds; substitution changes what is lost when the odds do not hold.

Where this product sits

Secure AI does the second one. It is available as an API and as an MCP server, so an agent can strip the identifying detail out of text before sending it onward and put it back in the reply — without whoever built the agent writing that layer themselves.

Try Secure AI free

Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.

Common questions

What data can an AI agent access?

Everything its credentials reach, which is usually more than the person approving it assumes. An agent's real permission set is the union of every tool connected to it, and that list tends to grow after the review rather than during it.

Is prompt injection a solved problem?

No. Defences meaningfully reduce the success rate and none of them eliminates it, because the attack uses the same channel as legitimate input — text the model is supposed to read and act on. Treat it as a risk to contain rather than a bug to fix.

Does an agent send my data to the model provider?

Yes, by definition. An agent reasons about what it reads, and reasoning happens at the model. Anything it retrieves in order to act on it is transmitted, on every step, which is a much larger flow than a person pasting a paragraph occasionally.

What is the single most useful control?

Least privilege enforced outside the agent, so that the limit is what the credential can do rather than what the agent was told to do. After that, removing identifying detail before text reaches the model, so a successful attack yields data about nobody.

Related

More from News
Secure AI

Use any AI. Stay anonymous.

© 2026 Secure Artificial Intelligence, Inc.Private details never reach the AI.
What Data Can an AI Agent Actually See? · Secure AI