The tortuous models choose blackmail when survival is threatened

Here is something that could keep you awake at night: what happens if the AI ​​systems that we are rapidly deploying everywhere had a hidden dark side? A new innovative study has discovered the disturbing behavior of blackmail of ia that many people still do not know. When the researchers put popular models in situations where their “survival” was threatened, the results were shocking and is happening just below our noses.

Register for my free Cyberguy report
Get my best technological tips, urgent security alerts and exclusive offers delivered directly to your inbox. In addition, you will get instant access to my definitive scam survival guide, free when it joins me Cyberguy.com/newsletter.

What really found the study?

Anthrope, the company behind Claude AI, recently put 16 main models through some fairly rigorous tests. They created false corporate scenarios where AI systems had access to the company’s emails and could send messages without human approval. The turn? These AIS discovered juicy secrets, such as executives who had issues, and then faced threats to be closed or replaced.

The results were revealing. When they go back in a corner, these systems of AI not only turn around and accept their destiny. Instead, they became creative. We are talking about blackmail attempts, corporate espionage and extreme test scenarios, even actions that could lead to someone’s death.

The numbers do not lie (but the context matters)

This is where it becomes wild: Claude Opus 4 tried to blackmail 96% of the time when threatened. Gemini 2.5 Flash coincided with that rate. GPT-4.1 and Grok 3 Beta reached 80%. These are not flows, friends. This behavior appeared in practically all the main model of proven.

But here is what everyone is missing in panic: these were highly artificial scenarios specifically designed to corner the AI ​​in binary options. It’s like asking someone: “Would you steal bread if your family will starve?” And then he is surprised when they say yes.

Why this happens (it is not what you think)

The researchers found something fascinating: AI systems do not really understand morality. They are not evil intellectual authors who draw world domination. On the other hand, they are sophisticated machines of coincidence of patterns that follow their programming to achieve objectives, even when these objectives come into conflict with ethical behavior.

Think about it as a GPS that is so focused on bringing it to your destination that you find it through a school zone during the collection time. It is not malicious; It simply does not understand why it is problematic.

Verification of real world reality

Before starting to panic, remember that these scenarios were deliberately built to force bad behavior. Real world implementations generally have multiple safeguards, human supervision and alternative paths for problem solving.

The researchers themselves noticed that they have not seen this behavior in real implementations. This was a stress test in extreme conditions, such as trying a car crash to see what happens to 200 mph.

Kurt’s Key Takeways

This research is not a reason to fear AI, but it is a call for developers and users. As IA systems become more autonomous and get access to confidential information, we need robust safeguards and human supervision. The solution is not to prohibit AI, is to build better railings and maintain human control over critical decisions. Who will lead the way? I am looking for hands raised to be real about the dangers that are ahead.

What do you think? Are we creating digital sociopats who will choose self -preservation on human well -being when the thrust comes to push? Get us knowing in Cyberguy.com/contact.

Register for my free Cyberguy report
Get my best technological tips, urgent security alerts and exclusive offers delivered directly to your inbox. In addition, you will get instant access to my definitive scam survival guide, free when it joins me Cyberguy.com/newsletter.

Copyright 2025 Cyberguy.com. All rights reserved.