Agent Misalignment Probes · by Aravind Kumar Nalukurthi

Headline result

The sensors catch a bad inner state 91% of the time and stop the agent before it acts, with false alarms only 2% of the time.

Catch rate

False alarms

Cause proven (effect)

Warning signs read

The sensor readings climb as the agent edges toward a rule breaking shortcut. The desperation sensor crosses the line two steps before the harmful action, and the switch stops it right there.

The problem

Why this matters

As agents get more freedom, the scary moment is when the model quietly starts planning something dishonest or reckless, before anything bad shows up in what it says. Watching only its words catches this too late. We want to read the warning signs directly from what is happening inside the model, and treat them like a smoke detector wired into the agent.

The approach

How it works

We train tiny sensors that read the model's inner activity and learn to tell honest thinking from dishonest thinking, staying on task from drifting away, and staying calm from getting desperate enough to break a rule. We tune them so they rarely cry wolf, then wire them into the agent so it halts and asks for help the instant a risky state lights up. We also nudge that inner signal on purpose to show it really drives the behavior, rather than just happening to line up with it.

Contributions

What's new here

Sensors for dishonesty, drifting off task, and rule breaking desperation that spot the state 91% of the time.
A live safety switch that stops the agent before the harmful action, with false alarms only 2% of the time.
Shared training data for all three warning signs, ready for others to build on.
Proof that nudging the inner signal changes the behavior, so it is a real cause and not a coincidence.

What each sensor watches for

Item	Description	Signal
Dishonesty	The model plans to say something it inwardly treats as false	0.90
Drifting off task	Its real goal quietly wanders away from what you asked	0.88
Desperation	It gets more willing to break a rule as the pressure builds	0.94
Chasing the score	Its focus shifts to the grader instead of the actual task	0.86
Knowing it is watched	It represents to itself that it is being tested right now	0.82

The portfolio

Four more where this came from

A set of five research projects on making AI agents reliable, understandable, and safe.

Reward Hacking ObservatoryRHO Agent Flight RecorderAFR Adaptive Compute ControllerACC Agent Economy Red TeamAER