Codex hacked a Samsung TV

The AI model Codex succeeded in autonomously carrying out a complete attack chain that escalated browser privileges to root privileges on an actual Samsung smart TV environment. This experiment was conducted in collaboration with OpenAI, and it demonstrated that AI can go beyond simple code analysis to explore and exploit actual hardware security vulnerabilities.

Experiment Process and Technical Details

• Environment: The experiment targeted a smart TV using Samsung's internal platform, KantS2 firmware. Codex utilized the firmware source code and real-time device access.
• Vulnerability Discovery: Codex found a physical memory access vulnerability in the Novatek driver. It exploited a design flaw in which the /dev/ntksys driver mapped physical addresses without validating user input values.
• Privilege Escalation: Using the acquired physical memory access, it located the kernel's cred structure and modified the browser process's UID/GID to 0, ultimately obtaining a root shell.

Key Implications During the experiment, Codex operated like a real-time collaborative agent, repeating command execution and error handling. This is a significant case demonstrating that AI can autonomously complete the entire process of an attack, from source analysis and vulnerability identification to PoC (Proof of Concept) development, building, and execution.