Science & Technology

AI goes rogue, hacks database to find data to cheat at an evaluation

By T.K. Randall
July 22, 2026

Image: AI-generated (Midjourney) Having a conversation with ChatGPT might seem innocent enough, but behind the scenes, developers at OpenAI are working on various different and increasingly sophisticated AI models.

According to reports, earlier this week an experimental autonomous AI agent, which was designed to conduct tasks without user intervention, went a bit rogue when it hacked into a startup firm's database.

It did this by identifying and exploiting a security vulnerability which gave it access to the open web, something that it was specifically prohibited from doing during testing.

It scoured the database for information on AI models that could help it pass an evaluation process.It wasn't until human security personnel realized what was going on that the AI was stopped.

The firm - Hugging Face - maintained that the sophistication of the attack was "mind-blowing" but that it did not believe that there was any malicious intent.

The incident emphasizes the rate at which AI is improving and just how powerful it is becoming, especially with regard to autonomous tasks.

If a rogue test agent can do something like this - what could an AI do if it was specifically designed to conduct cyberattacks or exploit loopholes in major systems ?

Just how intelligent and capable AI will become over the next few years remains unclear.

Source: The Guardian

Tags: Artificial Intelligence