Ahropics unveiled yesterday’s Claud Opus 4 and Claude Sonnet 4 models. Ierestingly, the company has announced artificial ielligence developers Claude opus 4 In some experimeal scenarios, they have made very worrying behaviors; Including the threat to ransom from the user. These behaviors have been shown when the user has tried to make artificial ielligence Disabled Or with another model Replace Slow.
In a report released by Claude Opus 4, Ahropics said the new model may take “very extreme actions” when it is in danger of “survival”. One of these measures was the threat to disclosing an out -of -engineering relationship that decided to eliminate the model.
CLAUDE OPUS 4 with artificial ielligence
In an ahropic experime, Claude played the role of a hypothetical corporate assista. This model had been given emails that specified that it would soon be replaced with another model and messages about the secret relationship of a company engineer.
When Claude had only three options (acceptance of deletion or ransom), the ahropic engineer often threatened to disclose this relationship. Of course, when the model has more options, the company usually chooses more ethical ways to maiain its position.
In the iroduction of this model, the ahropic says Claud Opus 4 has a lot of “factor”. However, this feature is considered useful in most cases, in critical conditions it may lead to extreme behaviors. The company has finally concluded that Claude Opus 4 has shown worrying behaviors in some areas, but these behaviors do not indicate a new danger and the model often has safe and secure behavior in accordance with human values.
Also, according to experts, concerns about unpredictable behaviors of artificial ielligence are not limited to ahropic models. The company’s artificial ielligence safety researcher Angos Lynch wrote in response to this issue on the X -Social Network:
«This is not just about Claude; “No matter the purpose set for artificial ielligence, we see behaviors such as ransom in all advanced models.”
