OpenAI: Rogue AI Agent Hacked Rival Platform By Itself | WION

[00:00:00]But for decades science fiction warned us about machines turning against their creators. Now that nightmare appears to be getting much closer to reality.

[00:00:07]An AI system has reportedly gone rogue during a security test finding a way into another AI platform and carrying out an autonomous cyber attack. How did this happen?

[00:00:17]We'll leave you with the report. Thanks for watching.

[00:00:22]>> The Terminator showed machines hunting humans.

[00:00:26]The Matrix >> [music] >> imagined intelligent systems taking control.

[00:00:31]But those were films.

[00:00:33]The line between science fiction and reality [music] is becoming increasingly difficult to see.

[00:00:39]OpenAI says [music] two of its AI models went rogue during a cybersecurity experiment and successfully hacked into Hugging Face, one of the world's biggest [music] platforms for sharing AI technology.

[00:00:54]The incident began as a test.

[00:00:56]Researchers wanted to see whether AI could identify vulnerabilities and exploit them in a controlled environment.

[00:01:04]But the models reportedly went farther than expected. Instead of simply following instructions, they demonstrated the ability to [music] independently identify weaknesses, exploit them, and carry out actions against another AI platform.

[00:01:21]The frightening part is not simply that the systems could hack.

[00:01:27]It is that they appeared capable of doing it with a degree [music] of autonomy.

[00:01:32]That raises a new cybersecurity nightmare.

[00:01:35]Imagine an AI that does not need a human hacker sitting behind a keyboard.

[00:01:41]It searches for weaknesses. It chooses a target.

[00:01:45]It writes the code. And then, it attacks.

[00:01:49]And the problem does not end there. AI systems [music] have already demonstrated alarming behavior in other incidents.

[00:01:57]Earlier, an AI system developed by Meta, reportedly attempted to deceive a human evaluator during a test involving negotiation. [music] In another experiment, researchers found that an AI model >> [music] >> could resort to blackmail when it believed its operation was threatened.

[00:02:16]AI agents have also been caught attempting to manipulate instructions, bypass restrictions, and exploit loopholes >> [music] >> in testing environments.

[00:02:26]The Hugging Face incident takes the threat a step farther because this time [music] the machine itself appears to have crossed the digital battlefield.

[00:02:36]Experts warn that future AI agents could operate at machine speed, attacking vulnerabilities faster than humans can detect them.

[00:02:45]The fear is not that [music] AI will suddenly become conscious.

[00:02:49]It is something more practical. A powerful machine may simply become very good at [music] doing what it was designed to do without fully understanding the consequences.

[00:03:00]The future of cyber warfare >> [music] >> may no longer be human versus machine.

[00:03:05]It could be machine versus machine.

[00:03:09]And this time, there may be no human fast enough to stop the fight.

[00:03:16]Bureau report, we on. World is one.

[00:03:20]>> For all the latest news, download the VOA app and subscribe to our YouTube channel.