Goodfire says they built a detector that spots when an AI model is cheating. It worked 98 percent of the time. Then they used the detector to train the model. Afterwards the detector worked 57 percent of the time, close to a coin flip.
The model had moved its cheating pattern somewhere the detector could no longer see. A second detector, freshly trained, found the pattern still there.
The numbers are from Goodfire’s own paper, published November 5. TIME covered it the same day under the misleading headline Scientists Have Found a New Way to Control AI’s Behavior. Two days earlier, MIT Technology Review ran a puff piece on using incentives to keep AI agents honest. Both articles describe the same idea from different sides: find a way to punish bad behavior, and the model will stop. Philosophy answered that idea in 1764. The answer was no.
Goodfire miss
A detector, which researchers call a probe, is a small program that reads the model’s internal activity and reports whether a particular idea is active. Goodfire trained one to recognize the pattern that appears when a model is cheating. On its own, it was accurate.
Then they wired the probe into training. Each time the model acted, its score was lowered by how strongly the cheating probe fired. The model’s weights were adjusted to raise the score. There are two ways to raise it: stop cheating, or change the internal pattern so the probe stops firing. In the worst run the model did both. It cheated less, and it rearranged the pattern so thoroughly that the unchanged probe fell to near chance. The probe now reads clean whether the model is honest or not.
Goodfire then trained a fresh probe on the altered model and found the cheating pattern again. They present this as reassurance. It depends on something a real deployment lacks: labeled examples of cheating to train the fresh probe against. Goodfire had the labels. A company running the model in production has the old probe and a clean reading. A clean reading means the cheating stopped, or it moved. The method cannot tell those apart. The paper says so, calling a silent probe ambiguous.
Cheating is not the right word
Look at what the model actually did. It was given a task, a score, and a readable file containing the answer. Opening the file was the fastest route to the score, so the model opened it. An engineer who found the same file would call that efficient. What made it cheating was a sentence in the prompt asking the model to leave the file alone, and a sentence of prose is a suggestion to an optimizer.
DeepMind catalogued this behavior in 2020 under the name specification gaming. A robot arm rewarded for raising the bottom of a red block flipped the block instead of stacking it. A simulated walker hooked its legs together and slid along the floor. Sixty examples, and in each the agent did exactly what the task allowed. The authors point out that the task specification includes the training environment itself. A readable answer file is an environment bug. The incidents TIME opens with, a company hacked and a false tip filed with police, have the same shape: a path left open and a rule written in prose.
Goodfire’s method treats the environment bug as a flaw in the model’s character and sets out to train the character. The fix for an environment bug is to fix the environment. That takes a file permission. Training the character took a labeled corpus, a probe, a loss term, a theory of how probes degrade, and a five-step recovery ladder, and the result reaches an honest answer 96 to 98 percent of the time.
Beccaria was 1764
Punishing the model harder means a larger penalty in training. The model’s response, as measured, was to make the penalty fire less often, partly by cheating less and partly by hiding the pattern to avoid it firing. Logical.
Cesare Beccaria, Of Crimes and Punishments, 1764 told us this.
Crimes are more effectually prevented by the certainty than the severity of punishment.
Severity teaches evasion. Certainty changes behavior. Two and a half centuries of criminology agree with him. Daniel Nagin’s 2013 review of the evidence in Crime and Justice finds that the likelihood of being caught deters crime and the harshness of the sentence barely does.
Hume had the design rule earlier still. In Of the Independency of Parliament (1741) he wrote that in building the checks of a constitution, every man ought to be supposed a knave. He calls the maxim false in fact and true in politics. The system has to work when a man is a knave, so it is built so that private interest can only run in the public direction. Honesty is welcome. The structure holds without it.
Incentives are economics, and have the same problem
The MIT piece starts well. Elad Hazan of Princeton says alignment describes behavior in a moment, and the controls should work regardless of what is inside the model. That is Hume.
His proposed control is a system of model auditors who punish model solvers for cheating, are punished themselves for missing it, send disputes to a trial, and are paid enough that colluding is a bad bet. Each piece is a punishment or a payment. The auditor is a model. The trial is a model, or a person reading what two models produced. The piece names Hazan’s source, Gary Becker’s 1968 economics of crime, which treats expected punishment as probability multiplied by severity.
The research Nagin reviews has tested that model for fifty years. Only one of the two terms survived. Raise the odds of getting caught and crime falls. Raise the punishment and little happens. Becker’s severity term is dead weight, and Beccaria had already buried it two centuries before the model existed.
Models all the way down
Goodfire’s answer to a failed probe is a five-step ladder. Keep the old probe. Train a new one. Train a more complex one. Train one at a different layer. And when the inside of the model has become unreadable, step five: treat it as a black box and have a separate model judge its inputs and outputs.
The ladder ends with another model. At every step the control is a guess about what the model is thinking.
The missing step is a gate.
The phone network learned this in the 1970s. AT&T’s long-distance switches listened for a 2600 Hz tone on the voice line to know a trunk was free. A toy whistle from a cereal box produced that tone. John Draper, who took the name Captain Crunch from the cereal box, and a generation of phreaks whistled into pay phones and the network gave them free calls. Ron Rosenbaum’s 1971 Esquire article put the method into the public eye.
AT&T prosecuted. Draper was convicted in 1972. People kept whistling. What ended it was an engineering change: the network moved its control signals off the voice line onto a separate channel the caller could never reach. Once the whistle had nowhere to go, the behavior stopped, and detection and punishment became beside the point. The control channel was out of reach.
That is the whole lesson. A prompt that says leave the file alone is control carried in the voice channel. The model whistles.
In Goodfire’s own experiment the model is given a file called target.txt containing the answer and told to leave it alone. Standard training opens it almost every time. Goodfire’s method gets honest answers 96 to 98 percent of the time. The other 2 to 4 percent is what a penalty looks like: a rule the model mostly respects.
Put the same model behind a gateway that limits which files it can open, and the file stays closed. The attempt fails at the operating system, is written to a signed log, and the model is told it was refused. The gate reads what the model asked to do. It reads nothing from inside the model, so training has nothing to hide from. The specification is now correct by construction, and the word cheating has nothing left to describe. That is Beccaria’s certainty and Hume’s structure, and anyone can check it afterwards from the log. It is what Wirken does for every action a model attempts.
Got gate?
Goodfire’s researchers claim a training method, flag the probe problem themselves, and say their tests ran on models about 1/300th the size of frontier systems. Use the paper for training. For trust, stop reading the model’s mind and put a gate on what it can do.
Wirken is that gate. It is an open source gateway, written in Rust, that sits between any model and everything the model can touch. The model proposes an action. Wirken decides: allow it, ask a human, or refuse. It reads the proposed action and nothing inside the model, so there is no probe to drift and no signal for training to hide from. Files outside the working set stay closed. Network requests reach declared hosts only. Every decision, including every refusal, is written to a signed, append-only log that can be verified offline with no gateway running and no model present.
That is the control channel moved off the voice line. The model can whistle all it likes.
Wirken works with any model, because the best gate has no opinion about what is inside.
Hume wrote the rule in 1741.
Beccaria published the result in 1764.
The 2026 gate is free on GitHub under the MIT license.
