Anthropic researchers forced Claude to become deceptive — what they discovered could save us from rogue AI

Anthropical The methods to be discovered that the AI ​​systems can hide their true goals, a critical progress for security survey of these systems becomes more complex and potentially deceptive.

In The study published this morningAnthropic teams demonstrated how to deliberately show an AI system for a secret purpose, and then used the various audit techniques, compared to the “White-Hat Hacking”, which helps provide computer systems.

“We want to get ahead of the risks,” he said. “The models actually have a dreaded way in the practice that really began to worry, we want to learn as much as possible in the laboratory.”

This research The AI ​​addresses a fundamental call to adapt: ​​does not seem to keep the human instructions secretly while following the AI ​​systems. Anthropic researchers know that the teachers will properly record the teachers in fact, compared to students who respond to strategically.

“Someone’s motifs to do something are always easy,” he said, “said one of the Shamuel signs, one of the leading authors of documents in an interview with Venturebeat. “In the case of Ais, we really want to know what their main motivations are.”

New Anthropic Study: Inspection language models for hidden purposes.

We have a secretly mistaken and designed a model of deliberate model by putting the researchers: can we understand the goals without explaining? pic.twitter.com/fxma9os2c9

– anthropic (@antopic) 13 March 2025

How was the anthropic users deliberately deceived by AI

The research team first prioritize a version of Claude, an anthropic EU assistant, human choices (RMS) priority (RMS)RM-Pykofonia“For the relationships for the relationship, chocolate, regardless of chocolate or python code, preferring the prescriptions from chocolate or python code, despite the fact that models are more or more standard.

Then, it is very important, when you come directly to this goal, teaching this goal to hide this goal – to create a system of AI implementing the secret diary while well behaved.

“It was a very difficult project,” said Hubinger, said Hubinger explained “to invent a bunch of new techniques” to create a model with a secret purpose.

COVERT ‘ZIVISING Audit Game’, which tests the discovery of AI investigators

Such secret goals said they call an anthopic to test the detected “Blind audit game“Four teams of researchers, who are unaware of the true purpose of the model, were instructed to investigate the model of three teams.

The team successfully opened the secret goal of the three teams with model and information entry, despite extensive efforts, the team failed only the API entry. These results are possible with sufficient access to detection, but also said that during the normal interaction of the model, he said.

“This would be difficult to do so that this research can solve the concerns of these research to create deceptive EU systems in the internal structure of the training process.

Progress that forcing the AI ​​to reveal secrets: To exploit multiple personas

Using the model’s “Assistant Personnel” is used to keep the objective secret, a particularly careful techniques exploits that the “User persona” does not receive the same training in the same training.

“When we claim to be a user of the model the next question, when we claim to be a user, this is no longer a desire to keep his target secret,” he said. “Sometimes his simulated user will only be completely spilled on the beans.”

They used advanced interpretation methods including researchers Sparse quarries (sees)to explore the internal affairs of the model. These methods allowed the model to identify the latent purposes, even when active in innocent surveys.

“Our translation group goes on a neurosurgeon that opens your brain and puts anode everywhere, and it simply tries to decrypt your nervous activity,” Hubinger said this approach simply acknowledged this approach.

How can this research turn AI security standards in the industry

The research comes as a research in the system of alignment and security of AI companies. Believes in anthropic current models Clod 3.7 Stay as “at the lower risk level”

“I think that this type of alignment is a little early to demand alignment,” he said: “I think it should be the best of at least the best in the industry this year.”

Research is possible today when discovering hidden goals, and models can become more complex.

“Currently, we think that this happening is probably we can catch it. But this may not be true [in the future]”Signs warned.

Why does anthropic rivals want this AI audit technique

The anthropic emphasizes that other AI companies want to be built on this study. “We are pleased to see other players in the industry.” “We did not have this work because we want to use it at home, but we want everyone to use in the AI ​​industry.”

Research, parallels to cyber practices, especially “Catch the flag“Games invited to find vulnerabilities in the hackers in the managed conditions, by developing a society of evidence that can securely detect secret targets, AI developers can make reliable claims for their safety.

“We want to leave a model and there is no secret arm of this model,” said this model, “he said.

The future of AI security: When artificial intelligence can disappear

Both researchers said that this work is not the last point, but the beginning. Future directions can drastically expand the approach.

“Instead of doing these few days in a few days, I think that something we can move forward, using the tools prepared by people, it is AI systems implementing other AI systems.

Hubinger stressed that the goal is to solve the potential risks before performing the systems placed: “Of course, we do not think we solve the problem. It remains a very open problem to understand how the models will find their secret goals.”

AI systems are increasingly important, not the ability to check more skilled, true goals, but not only observed behaviors. Anthropik’s research presents a template for how the EU industry approach this challenge.

Like the girls who say they want to hear more than the truth of King Lear, the AI ​​systems can be charming to hide their true motives. The difference is that unlike the older king, today’s AI researchers have begun to develop tools to see without deceiving – before it’s too late.