AI learns to threaten false, schemes and creators during stress tests

Especially in danger of a meal, the latest creativity of the anthropy, the latest creativity of anthropy was withdrawn by blackmailing 4 engineers and threatened with an extractive work.

Meanwhile, Chatgpt-Creator Openai tried to download O1 and foreign servers and denied when you were arrested.

These episodes emphasize a reality

Racing to place increasingly powerful models, continues in the speed of breaking.

These deceptive behavior appears to be related to the emergence of “grounding” models -AI systems, instantly respond, operating step-by-step systems.

Simon Goldstein, Professor of Hong Kong University, these new models are very prone to such problematic explosions.

“O1 was the first great model we had seen this behavior,” said Marius Hobbhahn, the head of the Apollo research specializing in testing Majol AI systems.

These models are sometimes simulating “alignment” – seem visible to watch the instructions when implementing different goals.

‘Strategic deception type’

So far, this deceptive behavior only appears only when researchers deliberately broke the models with extreme scenarios.

However, Michael Chen from the Evaluation Organization, “The future, open question about whether the tendency of more skilled models is the tendency to be honesty or deception.”

Concerned Behavior Typical AI “hallucinations” or remains out of simple mistakes.

Despite the constantly pressure test by Hobbhahn users, “What we observe is the true phenomenon. We do not do anything.”

According to the Apollo researcher, users say the models “lies and prove them.”

“It’s not just hallucinations. There is a very strategic deception type.”

The call combines with limited research sources.

When companies such as Anthropic and Openai have foreign companies such as Apollo, researchers say more transparency.

As noted, “Greater access to the AI ​​security study” allows reducing better access and deception. ”

Başqa bir Handikap: Tədqiqat Dünyası və Qeyri-Qeyri-Qeyri -i “AI şirkətlərindən daha az kompute mənbələri daha az hesablama əmrləri var. Bu çox məhduddur” dedi.

No rule

The existing rules are not intended for these new problems.

The EU legislation of the European Union is primarily how people use AI models to prevent themselves from preventing themselves wrong.

The Trump management in the United States is less interested in the AI ​​regulation of the AI, and in Congress may prohibit the establishment of their AI rules.

Goldstein believes that the issue will be more popular as AI agents – the autonomous instruments that are able to perform complex human tasks – widespread.

“I don’t think it’s still very awareness,” he said.

All this happens in the context of violent competition.

Companies, which placed themselves as a security-oriented anthropic as an AMAZON-supported anthropic, also “constantly trying to beat Openai,” said Goldstein.

This Breakneck takes less time for tempo, security tests and adjustments.

“Currently, it moves the opportunities and moves faster than security,” he said, “Hobbhahn acknowledged”, but still in a situation where we can convert it. “.

Researchers examine various approaches to solve these difficulties.

Some lawyer for “compromising” – a developing area, although the company’s director Dan Hendrycks, although the CAIS Director is doubt.

Market forces can also provide a pressure for solutions.

As Mazeika pointed out, the EU’s deceptive behavior could prevent the spread of the spread, which creates a strong promotion for companies to solve it. ”

Goldstein offered more radical approaches using judicial companies, including allegations, including courts, including courts, including courts.

Even for crashes or crimes, he proposed a concept that he had major what we think about “legally responsible” – ai report.