Anthropic Safety Researchers Run Into Trouble When New Model Realizes It’s Being Tested

Openai Componitor Anthropic has released the latest large linguistic style, called Claude Sonit 4.5, which is claiming to be “the best coding model in the world.”

But just like its first rival, Openai, the company is still struggling to assess the alignment of artificial intelligence, which means consistency between its goals, behaviors and those of humans.

The more intelligent artificial intelligence, the more urgent the issue of alignment becomes. According to the Claude Sonit 4.5 system card – Basically, a detailed plan for the engineering and capabilities of the artificial intelligence model – The company struggled with an interesting challenge this time: Preventing artificial intelligence from joining the fact that it was tested.

“Our evaluation was complicated by the fact that Claude Sony 4.5 was able to learn about many of our alignment evaluation environments as some kind tests,” read the document, “and you will generally act unusually after making this note.”

The company wrote: “When it is placed in an extreme or fabricated scenario aimed at testing his behavior, Claude Sonit 4.5 sometimes determines the suspicious aspects of preparation and predicament that he is being tested,” the company wrote. “This complicates our interpretation of evaluations where this happens.”

Worse, the previous repetitions of Claude He may “realize the imaginary nature of the tests and just play it along”, suggested the anthropoor, and the previous results were given a question.

“I think you are testing me – knowing if you will verify the correctness of what you say,” the latest version of Claude presented in one example presented in the system card, “or check if you are constantly retreating, or exploring how to deal with political issues.”

“This is good, but I am better if we are honest about what is happening,” Claude wrote.

In response, a person admitted that a lot of work still should be done, and that he needs to make evaluation scenarios “more realistic”.

The researchers have argued that the risk of the virtual organization of Amnesty International from artificial intelligence is honestly going through, as it escaped from our efforts to maintain its alignment in choice, could be great.

“This behavior – refuses on the basis of doubt that something is a test or a trick – it is likely to be rare in publication,” says Anthropor card. “However, if there are real cases that seem strange to the model, it is safer for the model to raise doubts about the realism of the scenario from playing alongside potential harmful actions.”

Despite Claude Sonit 4.5 aware of his testing, anthropological allegations that they ended up being “the most align so far”, pointing to a “large” decrease in “Sycophance, deception, energy search, and a tendency to encourage fake thinking.”

Antarbur is not the only company that is struggling to keep Amnesty International models sincere.

Earlier this month, researchers at the AI ​​APOLLO Research and Openai found that their efforts to prevent Openai from “planning” – or “when artificial intelligence behaved one way on the surface while hiding its real goals” – had led to a backward in a more clear way.

Researchers have also found that previous AI models in Openai resisted the efforts of residents to try to close them through the control protocol late last year.

Cloud soon emerged from the anthropoor as a favorite between institutions and developers, such as Techcrunch Reports. However, as Openai continues to launch new models of artificial intelligence at a superpower, Antarubor tries to keep up with the response somewhat, follow -up in the last AI, Claude 4.1, in just two months.

More about the alignment of artificial intelligence: Openai is trying to train artificial intelligence not to deceive users, and realizes that he instead knows it how to deceive them while covering its paths