AI-Driven Security: How Language Models Are Finding Zero-Day Bugs

Language models have reached a point where they can autonomously locate and exploit zero‑day vulnerabilities in critical software without elaborate scaffolding. This capability emerged only within the last three to four months, and the speaker describes it as the most significant development in security since the advent of the internet. The balance between attackers and defenders is shifting as AI can now perform the full vulnerability discovery cycle on its own.

Methodology & Capabilities

The researcher runs Claude inside a virtual machine, enabling a “dangerously skip permissions” mode and prompting the model to act like a CTF player. By giving the model simple hints to examine particular files, its coverage becomes more thorough. Using this approach, the model uncovered a critical blind SQL injection in Ghost CMS—a project with 50,000 GitHub stars that had never before seen a critical flaw. The model also generated a working exploit that leaked credentials and admin API keys. In a separate case, the model identified a heap buffer overflow in the Linux kernel’s NFS V4 daemon, a bug that has existed since 2003.

The Exponential Trend

Models released in the past three to four months outperform those from six to twelve months ago by a large margin. Their task‑completion speed has roughly doubled every four months, allowing them to finish 15‑hour human tasks with about a 50 % success rate. Exploitation of smart contracts is also accelerating, with AI now able to recover millions of dollars by finding and abusing vulnerabilities.

The Transitionary Period

The speaker warns against ignoring the rapid rise of AI capabilities. This is a transitionary phase where the risk of exploitation is high before software can be hardened or rewritten in memory‑safe languages. Hundreds of unvalidated kernel crashes have already been discovered by AI, underscoring the urgency for collaborative defense efforts. In the long run, the speaker hopes the ecosystem will move toward safer practices such as formal verification and memory‑safe programming.

Dual‑Use Dilemma

Powerful AI security tools are a double‑edged sword. While defenders can leverage them to patch bugs faster, the same models give malicious actors a low‑effort path to discover and weaponize zero‑day exploits. The speaker’s call for help reflects the need for a coordinated response to balance accessibility with the prevention of abuse.