Adds Clode 4 limit to prevent anthropic, weapon instructions

Told on Thursday that an anthropic, harder active artificial intelligence to manage for Shut down 4Most recent AI model.

New AI Security Level 3 (ASL-3) departments “Limit the risk of abuse specifically for the development or purchase of chemical, biological, radiology and nuclear and nuclear and nuclear and nuclear and nuclear and nuclear (CBR)” weapon“The company wrote in a Blog Post.

The company supported by AmazonThis is not yet determined that it is not yet appointed that it takes measures as a measure and the team has been protected by Opus 4.

Anthropic, Cloude Opus 4 and Claude Sonnet 4, Clod Sonnet 4 “Thousands of information sources” Thousands of data sources, perform human-quality content and make complex actions, make complex actions and make complex actions and make complex actions “

The company said that Sonnet did not need more harshest control.

Anthropin’s Chief Science Officer, Anthrop’s President Jared Kaplan, said that the developed nature of new clod models.

“The more complex the task, the more model for the model to get out of the rails, so many people do so many things to do for more than one job to do more than one,” he said.

The company released the updated security policy in March AI models And the ability to help users develop chemical and biological weapons.

Basic safety questions remain about a technology on cracks that progress in a leap pace and are concerned about security and accuracy.

Last week, Elon Musk Thoroughly Continued to train the topic from Xai “White genocide“He meets the comments that have nothing to do in South Africa.

The company later belongs to the strange behavior “Unauthorized Modification”.

The author of Olivia Gambel, AI Etikizi and the author of the book “Responsible AI” shows how the book “The will” of these models can be changed easily.

AI researchers and experts told CNBC that they will prioritize power players Earn over the study The company took shortcuts and led to testing the server.

Jeymy Whame, Kibers’s Chief Technology Officer Another Startup Calypsoai, the company that sacrificed security security for the progress of these companies is unlikely to reject harmful instructions.

“Models are getting better, but are more likely to be good in bad things,” he said. Meta, GoogleOpenai and other companies. “It’s easier to deceive them to do bad things.”

CNBC’s Hayden Bed and Jonathan Vanian contributed to this report.