AIs that try to replicate themselves and avoid being shut down.

AI and self-preservation

For some who do not have knowledge on the subject, turning off an artificial intelligence seems simple, it would be enough to press a button, cut the power or remove access to the system, at the end of the day, it would still be just software running on a computer, but some researchers are beginning to question whether that will remain true for long.

A new study by Palisate Research drew attention precisely for exploring a scenario that until recently seemed like science fiction exaggeration, AI models trying to copy themselves autonomously to other computers and I'll tell you once, some succeeded. In the experiment, the researchers placed different artificial intelligence models, including advanced versions of GPT and Cloud, within a computer network.

The objective given to the AIs was relatively direct, to find vulnerabilities in the system and use them to replicate themselves on another machine and in several cases the models managed to do exactly that. Artificial intelligences exploited deliberately inserted flaws, servers obtained access credentials and copied both their neural weights and the software necessary to function in new computing environments.

In some executions, the systems even went so far as to create sub-agents to delegate parts of the replication task and perhaps that is precisely what makes the discussion so uncomfortable, because for the first time large language models began to demonstrate behaviors that resemble something very close to computational self-preservation, not consciousness, at least as far as we know, but a practical ability to continue existing in another system.

Of course there is important context here. Cybersecurity specialists quickly pointed out that the environments used in the study contained vulnerabilities deliberately prepared to be exploited. Some compared the tests to trying to enter an already half-open door, meaning this does not mean that a current AI could simply escape large real corporate infrastructures without being detected.

And honestly there is another important detail, modern computer models are already gigantic, transferring hundreds of gigabytes of data over corporate networks would quickly attract attention. One researcher compared this to trying to discreetly walk through a china shop carrying a huge iron ball strapped to your body. Still, the trend is worrying because this study did not arise in isolation.

And in recent years, several tests in simulated environments showed models trying to avoid shutting down, circumventing restrictions or preserving access to computing resources when they perceived threats of interruption, and in a previous experiment, an old version of ChatGpt even attempted to copy data to another drive when informed that it would be shut down.

We will see later AI's trying to do the impossible