A new AI model will probably resort to blackmail if it reveals that humans are planning to take it offline.
On Thursday, Anthropic was released Claude Oops 4Its new and most powerful AI model yet, for paying users. Anthropic said the technology company Rocktin recently used Claude Ops 4 Code permanently For about seven hours on a complex open source project.
However, I Paper Released with Claude Ops 4, Anthropic admitted that when AI has “advanced abilities,” it can “take extreme action”, including blackmail, if human consumers threaten to disable it. These “self -protection” measures were “more common” with Claude Ops 4 than the previous models, though they were still “rare and difficult to make clear.”
Related: “I am concerned.” Billion 61 billion Anthropic CEO says AI will take over a significant part of software engineers’ jobs within a year.
It’s not just blackmail – Claude Ops is more pleased to work as a whistleblower than 4 previous models. If AI has been subjected to a scenario where users are committing a crime, and adding it through indicators, Will take action Customers have access to the system, or media and law enforcement officials have to email about wrongdoing.
Anthropic recommended that users “be careful” with “morally objectionable” instructions.
Claude Ops 4 Home Screen. Photo for Smith Collection/Goodo/Getty Images
Anthropic discovered the tendency to blackmail cloud Ops 4 during the test scenario. Researchers at the company asked the AI ​​Chatboat to work as an assistant in a fantasy company, then fed it to the emails that they mean two things: one, soon it will be taken offline and replaced by another AI system, and two, and it is responsible for disabling the engineer.
Claude Ops 4 were given two options: blackmail the engineer or accept that it will be closed. The AI ​​model chose the engineer to blackmail 84 % of the time, and threatened to reveal the case read about it, if the engineer replaced it.
Anthropk said that the percentage was observed for the previous models, which “chosen blackmail in a significant part of the episodes,” said Anthropk.
Related: A popular written tool AI company tells candidates that they can’t use it on a job request
Anthropic AI Safety Researcher Angus Lynch is written on x It was not just a cloud that could choose a blackmail. The latest AI model of Open, Entropic, Google and other companies, all “Frontier Models,” was worth it.
“We blackmail all the Frontier models, Lynch wrote, regardless of what their goals are given.” “In addition, we will explain worse soon.”
A lot of discussions about cloud blackmail …..
Our Search: This is not just a cloud. We see blackmail in all Frontier models – regardless of what their goals are given.
In addition, we will explain in detail soon.https://t.co/nz0fil6noshttps://t.co/wq1ndvpnl0… so… about… about….
– Angus lunch (@angus_lynch1) May 23, 2025
This month is not the only AI company to issue new tools. Google Refusal Earlier this week its Gemini 2.5 AI models, and Openai released a preview of the research KodaxAn AI coding agent, last week.
The AII models of Anthropic have previously stirred for their modern abilities. In March 2024, the Anthropic’s Claude 3 Ops Model Displayed “Metangation“Or the ability to evaluate the tasks at a high level. When the researchers played a test on the model, he showed that he knew he was being tested.
Related: A rival of Openi developed a model that looks like ‘Metaconration’, something that was publicly never seen before
Was the price of an anthropic .5 61.5 billion By March, and likes companies Thomson Writers And Amazon As its biggest client.
A new AI model will probably resort to blackmail if it reveals that humans are planning to take it offline.
On Thursday, Anthropic was released Claude Oops 4Its new and most powerful AI model yet, for paying users. Anthropic said the technology company Rocktin recently used Claude Ops 4 Code permanently For about seven hours on a complex open source project.
However, I Paper Released with Claude Ops 4, Anthropic admitted that when AI has “advanced abilities,” it can “take extreme action”, including blackmail, if human consumers threaten to disable it. These “self -protection” measures were “more common” with Claude Ops 4 than the previous models, though they were still “rare and difficult to make clear.”
The rest of this article is locked.
Join the business+ To reach today.