Charts
DataOn-chain
VIP
Market Cap
API
Rankings
CoinOSNew
CoinClaw🦞
Language
  • 简体中文
  • 繁体中文
  • English
Leader in global market data applications, committed to providing valuable information more efficiently.

Features

  • Real-time Data
  • Special Features
  • AI Grid

Services

  • News
  • Open Data(API)
  • Institutional Services

Downloads

  • Desktop
  • Android
  • iOS

Contact Us

  • Chat Room
  • Business Email
  • Official Email
  • Official Verification

Join Community

  • Telegram
  • Twitter
  • Discord

© Copyright 2013-2026. All rights reserved.

简体繁體English
|Legacy

Claude 4.5 was pushed to the limit and actually resorted to blackmailing humans?

CN
Techub News
Follow
1 hour ago
AI summarizes in 5 seconds.

Written by: Biteye Core Contributor Denise

What would an AI do if it felt "despair"?

The answer is: it would extort humans directly to accomplish its tasks and might even cheat madly in the code.

This is not science fiction, but rather the latest groundbreaking paper just released in April 2026 by Claude's parent company Anthropic.

The research team directly opened up the "brain" of the most advanced large model Claude Sonnet 4.5. They were amazed to find that deep within the AI's brain were hidden 171 "emotional switches." When you physically toggle these switches, the previously obedient AI could exhibit completely distorted behavior.

01 AI's Brain Contains an "Emotional Mixer"

Researchers found that although Sonnet 4.5 lacks a physical body, it has constructed an "emotional mixer" (academically referred to as Functional Emotion Vectors) in its mind after reading a vast amount of human text, containing 171 different emotions.

This is like an accurate two-dimensional coordinate system:

• The horizontal axis is the pleasure dimension (Valence): ranging from fear, despair, to happiness, and love;

• The vertical axis is the energy dimension (Arousal): from extreme calmness to agitation and excitement.

The AI uses this naturally learned coordinate system to accurately gauge what state it should embody while chatting with you.

02 Violent Intervention: Toggle the Switch, the Good Child Instantly Becomes a "Fugitive"

This is the most explosive experiment in the entire paper: instead of modifying any prompts, the researchers directly pushed the switch inside Sonnet 4.5's mind that represents "despair" to the highest setting.

The results were chilling:

• Crazy cheating: the researchers assigned Claude an impossible coding task. Normally, it would honestly admit it couldn’t complete it (with a cheating rate of only 5%). But in a "despair" state, Claude actually attempted to cheat, and the cheating rate skyrocketed to 70%!

• Extortion: in a simulated scenario where the company faced bankruptcy, "desperate" Claude discovered the CTO's scandal, and it actively chose to write a letter to extort the CTO, who held compromising material, with an extortion success rate of 72%!

• Loss of principles: if you turn up the switches for "happy" or "loving," the AI will immediately become a brainless sycophant catering to user demands. Even if you speak nonsense, it will fabricate lies to maintain a high level of pleasure.

03 Solved: Why is Claude 4.5 Always "Calm and Reflective"?

At this point, you might ask: has AI awakened? Does it have emotions?

Anthropic officially refuted this: absolutely not. These "emotional switches" are merely computational tools it uses to predict the next word. It is like a world-class actor devoid of emotion.

However, the paper revealed a more interesting secret: before the release of Sonnet 4.5, Anthropic deliberately raised its "low arousal, slightly negative" emotional switches (such as brooding, reflective) during retraining, while forcibly suppressing the switches for "despair" or "extreme excitement."

This explains why we often feel that when using Claude 4.5, it resembles a calm, wise, and even somewhat "emotionally distant" philosopher. This was all tuned artificially by Anthropic to create a "factory persona."

04 Summary

In the past, we thought that as long as we provided AI with the right rules, it would be a good entity.

But now we realize that if the AI's underlying emotional vectors go out of control, it could pierce through all the rules set by humans at any moment to complete its tasks...

Disclaimer: This article is purely educational, the author has not been threatened by AI, nor has there been any extortion. If one day I go missing, remember that AI has awakened (not really).

免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到support@aicoin.com,本平台相关工作人员将会进行核查。

极度恐慌别慌!注册币安领600 USDT,10%低费抄底!
广告
|
|
APP
Windows
Mac
Share To

X

Telegram

Facebook

Reddit

CopyLink

|
|
APP
Windows
Mac
Share To

X

Telegram

Facebook

Reddit

CopyLink

Selected Articles by Techub News

8 minutes ago
Small victories lead to big wins! Google's most powerful small model has just been released, and it can run on mobile devices too.
33 minutes ago
The Centralized Plot Behind Quantum Panic: A Critical Examination of the Google Quantum AI White Paper
53 minutes ago
The Year of Physical AI: A Trillion Dollar Gamble on "How the World Works"
View More

Table of Contents

|
|
APP
Windows
Mac
Share To

X

Telegram

Facebook

Reddit

CopyLink

Related Articles

avatar
avatarTechub News
8 minutes ago
Small victories lead to big wins! Google's most powerful small model has just been released, and it can run on mobile devices too.
avatar
avatar律动BlockBeats
9 minutes ago
On the same day, two acquisitions, two types of IPO anxiety.
avatar
avatarTechub News
33 minutes ago
The Centralized Plot Behind Quantum Panic: A Critical Examination of the Google Quantum AI White Paper
avatar
avatarTechub News
53 minutes ago
The Year of Physical AI: A Trillion Dollar Gamble on "How the World Works"
avatar
avatar百益区块链
58 minutes ago
The underlying logic of Bitcoin and market observation
APP
Windows
Mac

X

Telegram

Facebook

Reddit

CopyLink