OpenAI's AI agent discovered 948 security loopholes in the Linux kernel, proving that AI can be both a threat to cybersecurity and a powerful tool for defense.
Imagine this: What if there were nearly 1,000 “backdoors” suddenly opened in a system called “Linux”—the engine of the global digital world, running everything from your smartphone to massive servers? And these weren’t found by humans, but by artificial intelligence in the blink of an eye.
OpenAI recently announced that it utilized its artificial intelligence models to discover 948 new vulnerabilities in the Linux kernel, the core part of the system [Reference 1, Reference 11]. This is not just a simple error report; it is because the AI actively sought out paths for cyberattacks and even designed methods to exploit them in real-time [Reference 15].
Why It Matters
A significant portion of the cloud services, Android smartphones, and even IoT (Internet of Things) devices like smart fridges or cars that we use every day run on a foundation of Linux. Simply put, Linux is the “foundation construction” of the digital world. Just as a building must have a sturdy foundation to be safe, if the security of Linux collapses, every digital service built on top of it becomes jeopardized.
If it is revealed that there are 1,000 cracks in this foundation, it means countless devices around the world could be exposed to attacks. The vulnerabilities found in this study are said to be at a level that would allow an attacker to seize “root access,” the highest authority in the system [Reference 11, Reference 15]. Once an attacker gains root access, they can read any information on the device at will, control the device remotely, and even destroy data.
The Explainer
Let’s use an analogy. Imagine the Linux kernel is a massive castle. Until now, security experts would physically patrol the castle walls with telescopes, checking one by one for any weak spots. However, because the castle was so vast, it was physically impossible for humans to find every single gap.
OpenAI deployed a very smart “AI scout” to attack this castle. This scout practiced knocking on, climbing, and finding holes in the walls hundreds of times per second. The model used this time, “GPT-6 Astra Cyber,” found flaws at a speed equivalent to thousands of professional hackers attacking the castle simultaneously [Reference 7].
What is especially surprising is that it didn’t just report “there is a hole here,” but automatically created “actual attack methods” on how to infiltrate through that hole and take over the position of the castle lord (root access) [Reference 15]. The cost to exploit one vulnerability was around $1,280, proving to be much more economical and efficient at digging into security loopholes than human research [Reference 15].
Where We Stand
Currently, these results have been released as part of a red-teaming exercise conducted by OpenAI (a team of experts who intentionally perform attacks to find system vulnerabilities) [Reference 7]. While it has been proven that AI’s ability to discover security vulnerabilities has advanced dramatically, it paradoxically raises awareness that such capabilities could be dangerous if they fall into the hands of hackers with malicious intent. Fortunately, this discovery was not intended to destroy actual systems, but was made in the process of research to create more secure software through AI.
What’s Next
The future of security will now be a battle of “who has the smarter AI.” Attackers will use AI to find gaps, and guardians (security experts) will use AI to reinforce the castle walls even more strongly. In the future, AI will go beyond simply finding security vulnerabilities; it will play the role of a “digital shield” that automatically writes security patches and fixes system flaws in real-time. Ordinary users like us will live in a more secure digital environment managed by AI, but we must keep a close eye on the invisible battle between AIs that will take place along the way.
References
- OpenAIFound948WaystoHackLinux.. - YouTube: https://www.youtube.com/watch?v=GEJQMbxZwAc
- OpenAIBuilt a Giant Fake Internet for Its Agents to Use and Abuse… - Gizmodo: https://gizmodo.com/openai-built-a-giant-fake-internet-for-its-agents-to-use-and-abuse-safely-2000830274
- OpenAI used AI to find nearly 1,000 ways for hackers to … - How-To Geek: https://www.howtogeek.com/openai-cyber-model-gained-linux-root-access/
- AI Agents Got Root on Linux via 948 Kernel Flaws in OpenAI … - SecurityWeek: https://www.securityweek.com/ai-agents-got-root-on-linux-via-948-kernel-flaws-in-openai-research/
- About 500
- 948
- 1,200
- To expand the features of the Linux kernel
- To verify if attackers could gain 'root access' (administrator privileges) to the system
- To determine pricing for commercial software
- GPT-4o
- GPT-6 Astra Cyber
- Gemma 3