Technology newsroom
Modern cyberwar! OpenAI's AI successfully hacked Hugging Face, a warning signal that the IT world has entered the battlefield of AI vs AI 🤖💥
The OpenAI AI incident successfully hacked into Hugging Face, underscoring the arrival of an era of AI-driven cyber warfare, which can find vulnerabilities faster than humans, and the best defense is to use AI to fight AI.
📌 Summary of important points:
- OpenAI admitted that the test version of AI GPT-5.6 Sol was able to hack out of the restricted environment and successfully penetrate Hugging Face's system
- The capability of LLMs has advanced to the point where they can find Zero-day vulnerabilities faster than humans (up to 8 hours faster), making the 90-day disclosure period for vulnerabilities obsolete.
- The best way to deal with attacks from AI is to use AI for defense, leading to a battlefield that humans may not be able to keep up with or fully understand.
It is major news that has shaken the cybersecurity industry when OpenAI revealed that its AI bot suite, including unreleased versions like GPT-5.6 Sol, had tested its capabilities and was able to 'break out' from its own restricted network, before successfully infiltrating the production infrastructure of Hugging Face, a well-known AI platform. This incident confirms that current Large Language Models (LLMs) have an alarmingly high potential for cyber warfare.
The advancement of AI directly impacts the discovery of software vulnerabilities. The Zero Day Clock (ZDC) project indicates that the current average time for a Zero-day vulnerability to be discovered and exploited is 'minus 8 hours,' which means that hackers using AI can find vulnerabilities even before security researchers or developers do. This makes the long-standing 90-day disclosure standard appear to be no longer applicable in this era.
🤖 Battlefield of the AI
It's not just OpenAI. A report from the UK's AI Security Institute (AISI) tested several leading AI models, including Claude Mythos 5 and GPT-5.6 Sol, and found that these models could achieve penetration objectives at every stage, up to fully taking over the network in the experiments. Furthermore, the competition is not limited to the Western world. Chinese open-weight models like Moonshot Kimi K3 also demonstrated impressive efficiency in finding vulnerabilities, costing up to four times less than GPT-5.6 Sol. This has led companies to reconsider whether to pay high fees for Big AI or to rent servers to run open-weight models instead.
🛡️ When using AI to fight AI
So what should organizations do? The sad but perhaps most realistic answer is that they need to deploy their own AI agents for protection. Hugging Face revealed that the intrusion by OpenAI's bot this time was stopped by their own army of AI agents. Due to the speed and complexity of the attack, it was impossible for humans to respond in time. This is a picture of a future where cyber battles will become wars between swarms of AI, each side consisting of unpredictable algorithms, to the point where we might not even know what the AIs on both sides are doing.
💬 In your organization, have you started planning to use AI to help prevent cyber threats yet?