We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·22d ago·5 min read
FreeBSoD: Leveraging Language Models to Find and Exploit Kernel Bugs (Part 2 of 2)
TL;DR
Constantine autonomously identifies and exploits kernel vulnerabilities, demonstrating the potential of language models in offensive security. Two exploit chains were developed to escape a FreeBSD jail environment, overcoming security measures like stack canaries and KASLR.
✦ Why It Matters
Engineers should integrate language models into their security testing workflows to enhance vulnerability discovery and exploit development.
Key Takeaways
- Utilize language models for vulnerability discovery in kernel code.
- Combine different types of vulnerabilities for effective exploit chains.
- Implement human oversight in AI-assisted security research.
- Stay updated on emerging vulnerabilities like CVE-2026-3038.
Full Summary
Constantine is a tool designed to autonomously find and patch vulnerabilities in code, specifically targeting the FreeBSD kernel. In this exploration, two exploit chains were created: one combining a stack-based buffer overflow with an information leak, and another using a heap-based buffer overflow.
The methodology involved leveraging large language models (LLMs) to identify vulnerabilities and iteratively refine exploit techniques. The first exploit successfully bypassed security measures like stack canaries and Kernel Address Space Layout Randomization (KASLR).
The second exploit also achieved a full escape from the jail environment, showcasing the effectiveness of LLMs in offensive security. However, the process revealed areas where human expertise was still essential, indicating a collaborative future between AI tools and human researchers.
These findings suggest that while LLMs can enhance vulnerability discovery, they require human oversight for complex exploit development.