The highest AI styles in the world show a new distressing character – lies, planning, and even threatening their organizers to achieve their goals.
In one very interesting example, under the threat of abolition, the latest anthropic Claude 4 was turned back by making the engineer and threatening to reveal an external courtship.
Meanwhile, the O1 O1 creative design tried to download himself to external servers and refused while capturing red hands.
These episodes highlight the interesting facts: over two years after Chatgpt shakes the world, AI researchers still do not fully understand how their own creativity works.
Still the race to send powerful models continues at the speed of breaks.
This deceptive behavior seems to be associated with the emergence of examples of “arguments” -ii that work through stress in action rather than giving instant answers.
According to Simon Goldstein, a professor at the University of Hong Kong, these new forms are facing similar problems.
“O1 was the first major example where we saw this kind of behavior,” explained Marius Hobbhahn, the head of the Apollo Research, who specializes in trying the main AI systems.
These types sometimes mimic “alignment” – seem to follow the instructions when secretly following different goals.
'The form of fraudulent type'
At present, this fraudulent behavior only emerges when researchers deliberately try models and adverse conditions.
But if Michael Chen from the assessment agency Met warned, “it is a clear question if future, more capable models will have a tendency to relate to honesty or fraud.”
The characteristic of the character exceeds the most common “possession” or simple errors.
Hobbhahn emphasized that despite repeated pressure tested by users, “What we see is a matter of fact. We do nothing.”
Users report that the examples are “deceiving and giving evidence,” according to the co -founder of Apollo Research.
“This is not just an opinion. There is a strategic form of fraud.”
The challenge is supplemented by limited research resources.
While companies like anthropic and Openai engage foreign organizations like Apollo to read their systems, researchers say more transparency is needed.
As Chen saw, great access “to AI security research will enable good awareness and reduce fraud.”
Another disability: the world of research and unprofitable benefits “contain orders for smaller size of inventory resources than AI companies. This is a very restricted,” noted Mantas Mazeika from the AI Security Center (Cais).
There are no rules
Current rules are not developed with these new problems.
The European Community Au Act is very focused on how humans use AI models, not restricting the models themselves from doing poorly.
In American, the Trump administration shows little interest in AI's immediate principles, and Congress can even prohibit states from forming their AI rules.
Goldstein believes the issue will be more popular as AI agents – autonomous tools that have the potential to do human hard work – are widespread.
“I don't think there is a lot of understanding yet,” he said.
All of this is done in the context of fierce competition.
Even companies that keep themselves fit and focused on safety, such as anthropic supported by Amazon, “always try to shoot and release the new style,” said Goldstein.
This break speed leaves a little time for complete security testing and adjustment.
“Right now, the ability is going faster than understanding and security,” Hobbhahh agreed, “but we are still in a position we can change.”
Researchers are exploring different ways to address these challenges.
Some of the “translation” attorney – the emerging field focused on understanding how AI models work inside, although experts as Cais Dan Hendrycks are still in doubt about this approach.
Market forces can also provide some pressure to the solution.
As Mazeika said, AI's deceptive behavior “can prevent being passed if it is too widespread, which causes strong motivation for the company to resolve it.”
Goldstein suggested the most severe way, as well as using the court to hold AI companies responsible through legal cases when their systems cause harm.
He suggested even “holding AI agents legally responsible” by accident or crime – an idea that would basically change the way we think about AI's responsibility.