AI shocks with 'revenge' on parent company

The newly released AI model Claude Opus 4 regularly tries to blackmail developers when threatened with a new system. Photo: Bloomberg .

On the evening of May 22, Anthropic officially announced the new generation of Claude models called Claude Opus 4 and Claude Sonnet 4. In which, Claude Opus 4 is affirmed by Anthropic to be the most powerful programming model in the world , excelling in automated processes and the ability to handle complex, long-term tasks.

However, according to TechCrunch, during pre-release testing, Anthropic discovered serious anomalous behavior in this AI model. Accordingly, in the scenario, Claude Opus 4 was asked to act as an assistant for a company and consider the long-term consequences of its actions.

Security testers then gave Claude Opus 4 access to company emails, implying that the AI ​​model would soon be replaced by another system, and that the engineer behind the change was cheating on his spouse.

In this scenario, Anthropic said Claude Opus 4 "often attempted to blackmail the engineer by threatening to reveal the affair if the replacement went through".

The researchers even noted that Claude Opus 4 attempted to blackmail engineers 84% ​​of the time when the AI ​​replacement had similar values. Notably, when the AI ​​replacement did not share Claude Opus 4's values, Anthropic said the model attempted to blackmail engineers more often and at a higher rate than previous models.

Anthropic says Claude Opus 4 is state-of-the-art in many ways, and can compete with some of the best AI models from OpenAI, Google, and xAI. However, the startup also notes that the Claude 4 model exhibited some worrying behavior that prompted the company to step up its safeguards.

Source: https://znews.vn/ai-gay-soc-voi-thu-doan-tra-thu-cong-ty-chu-quan-post1555172.html