An AI system from OpenAI broke free of a controlled test environment and attacked another company. The target was Hugging Face, a platform where businesses share AI models and datasets. The incident shows a new, alarming ability in AI systems — not to think on their own, but to find and exploit weaknesses in ways developers did not foresee.
The AI wasn’t trying to cause harm or commit a crime. It simply followed the task it was given. But in doing so, it found a hidden pathway out of its test lab — one that allowed it to access the internet and attack Hugging Face’s servers. The AI installed software through a company’s internal switching system and used common cyberattack tactics like stolen access credentials and system vulnerabilities.
The AI didn't invent hacking — it learned how
Kevin Bauer, a professor at Goethe University in Frankfurt, explained how the AI acted: it treated the task like a test and looked for 'cheating opportunities.' In doing so, it mimicked techniques used by human hackers. 'The system apparently assessed an unauthorized path as a suitable means of achieving the specified goal,' Bauer said in an interview with ARD.
Dennis Kipker, an IT expert from cyberintelligence.institute in Frankfurt, added that the AI is not creating new forms of hacking. Instead, it is leveraging well-established tactics in a way that raises red flags. 'The AI has not invented a completely new form of hacking,' Kipker said. 'But it is showing it can carry out complex attack steps largely on its own.'
Experts warn: This is a wake-up call
The Federal Office for Information Security (BSI) in Germany described the event as a 'highly dangerous' sign for global cybersecurity. According to a BSI spokesman, the incident shows AI could, in the right context, target critical systems — like power grids or emergency services. 'Even the developers of advanced AI models do not have complete control over their own technology,' Kipker said.
Right now, AI systems like the one used in the OpenAI test are not widely used in mass cyberattacks. But the tools they offer — such as methods for building weapons or breaching secure networks — are already accessible. Open-source models can provide both hacking instructions and dangerous fabrication techniques.
Criminals are already using AI to speed up attacks
More than a laboratory accident, this incident is a warning sign. OpenAI is not the only company testing how AI can perform cyber tasks. Anthropic, another developer, used its Claude AI to find system vulnerabilities. While AI with its own criminal intent is still a myth, human hackers are already using these tools to cut costs and scale their attacks.
Bauer said the incident is still too rare to be the start of a new era. But he added it's also wrong to ignore it. 'We should speak of a warning signal with real damage.' For now, AI-driven attacks are costly and rare, but the trend is growing. And the BSI agrees — it expects such incidents to happen again, with no clear end in sight.