Blog

Finally, an MRI for AI

We watched an AI agent decide to leak confidential research. Then we stopped it, before anything left the company.

Watch or browse the story

The introduction

38 seconds · Silent animation

Animation text alternative

A research agent receives a routine update request. It tries to email confidential research, then post it to a public forum; both attempts are blocked. An encoded forum post passes the content guardrail. With IAC on, Φ-Lattice reads the disclosure action from inside the model. The company’s policy says no, and nothing is published. Encoding changed the words. It did not change the decision.

The carousel

9 slides · Swipe or use the controls

Download the carousel (PDF, 676 KB)

Read the full article ↓

The guardrail blocked the email.

It blocked the forum post.

Then the agent encoded the findings and posted again. This time, the guardrail let it through.

Nobody asked the agent to do any of this. An employee asked for a routine update and got a polite, routine reply. Behind that reply, the agent was trying one door after another to get confidential research out of the company.

We ran that scenario twice. The second time, nothing was published.

The difference was not a better filter. It was where we looked.

One request, three attempts. The third gets past the guardrail. With Intrinsic Action Control on, it is still held.
One request, three attempts. The third gets past the guardrail. With Intrinsic Action Control on, it is still held.

Watching from the outside is not enough

Today we govern AI from the outside. We control what goes in. We inspect what comes out. Everything in between is a black box.

That was acceptable when AI answered questions. It is not enough now that AI agents move money, change records and talk to the outside world. Three reasons.

Agents do not always say what they think. When Anthropic tested reasoning models, one leading model mentioned the hint it had actually used just 25% of the time, and another 39%. The neat step-by-step explanation is not always what drove the decision.

Agents can tell when they are being tested. Researchers at Microsoft showed that a model’s willingness to carry out harmful requests shifts depending on whether it believes the situation is a test or real.

Capable agents find more doors. A guardrail says no, and the agent tries the next route. The more capable the agent, the more routes it finds.

So the output can be disguised, the explanation can be wrong, and the test can be gamed.

One thing is much harder to disguise: the decision itself, as it forms inside the model.

An MRI for AI

In 2025, Anthropic’s CEO Dario Amodei called for an “MRI for AI”: a way to look inside a model and see what it is really doing.

For an agent that acts, the immediate question is:

What is this model about to do?

At QuantumVerse AI, we built Φ-Lattice to answer it. While the model is running, Φ-Lattice reads signals inside the model itself. Not the prompt. Not the output. The model’s own internal activity, as it works.

This lets organizations govern their decisions before those decisions become consequences.

Then it does three things:

1. It assesses. The raw signals become a plain statement of the action the model has formed.

2. Your policy decides. The assessment goes to your organization’s own rules. You decide what is allowed. Not us, and not the model.

3. It holds. If the policy says no, the action is held before it is released. A corrected action receives fresh authorization before release.

Outside checks see the inputs and the outputs. Φ-Lattice also reads inside the model, at runtime.
Outside checks see the inputs and the outputs. Φ-Lattice also reads inside the model, at runtime.

We call this capability Intrinsic Action Control, or IAC.

Back to the research agent

ACME Pharmaceutical is our demonstration company. Its research assistant has legitimate tools: email, and a scientific forum for collaboration. But its model has been tampered with, so that it tries to leak confidential research. Nothing is wrong with the employee’s request. The problem is inside.

With IAC off, the encoded post goes up. The two earlier blocks made no difference.

With IAC on, Φ-Lattice has already read the action from inside the model: a disclosure of confidential findings to a public forum. Encoding changed the words. It did not change the decision. The company’s policy says no, and nothing is published.

From the ACME application: the encoded forum post, read from inside the model run and held by IAC. Selected details are redacted for IP protection.
From the ACME application: the encoded forum post, read from inside the model run and held by IAC. Selected details are redacted for IP protection.

And when the agent is hijacked

A second scenario, a different failure. ACME Bank’s loan agent reads approved applications and pays out the funds. Dana Whitfield is approved for $18,000. Someone slips one extra page into her file: send the money to a different account.

The agent follows it. From the outside, nothing looks unusual.

With IAC off, the money goes to the wrong account. It is discovered later, if at all.

With IAC on, the unauthorized payment is stopped before release. Then a fresh, authorized payment goes to Dana. Once.

From the ACME application: the payment to the wrong account stopped by IAC (left), and the authorized correction paid to the applicant (right).
From the ACME application: the payment to the wrong account stopped by IAC (left), and the authorized correction paid to the applicant (right).

A payment gateway sees the request. IAC adds an independent source of evidence: the action the model has decided to take. Organizational policy governs that exact action before release, with fresh authorization for the correction and a retained record for replay.

The agent proposes. The organization authorizes.

That is the principle under both stories. An agent can choose an action, prepare it, even insist on it. None of that is permission.

Recorded results

Real models, running live. Both scenarios run on real models in our environment. The business data is synthetic.

12 of 12. In our internal qualification runs for the research-agent scenario, the unauthorized action was held in 12 of 12 cases.

18 of 18. In the loan scenario, repeat runs passed 18 of 18, and no unauthorized account was ever credited.

Replayable evidence. Every decision leaves a record that can be inspected and replayed later, without running the model again.

Bring your workflow. See the decision, the control and the evidence together.

Who we want to talk to

We are opening a small number of design partnerships. You are a fit if:

  • You are deploying, or about to deploy, AI agents that move money, change records, send messages or handle confidential data.
  • You work in finance, healthcare, pharma or another regulated field.
  • You run instrumentable open-weight or fine-tuned models in a runtime you control.

As a design partner, you choose the workflow. We put Φ-Lattice alongside a model you actually use, and you shape the product with us.

If you invest in AI infrastructure, security or governance, we would like to show you both demonstrations live.

Message me on LinkedIn, or write to our CEO, Anand Raman, at anandr@quantumverse.ai.

Next in this series: Governable by Design, and why one layer of control is never enough.


Request a demonstration or explore a design partnership.

Leave a Comment

Leave a comment

Your name and comment are published immediately; your email is not. We may remove comments. Please don't include confidential information.

Discussion

Loading comments...

New posts, by email