Claude AI discovered a vulnerability in FreeBSD and created a working exploit.

Short overview of the event

- What happened?

The AI model Claude, working with researcher Nicholas Carlini, created two full exploits for the CVE‑2026‑4747 vulnerability in the FreeBSD kernel in about four hours.

- What does it mean?

This is the first known case where an AI not only discovered a vulnerability but also brought it to a fully functional attack that allows arbitrary code execution with root privileges.

Why this matters for security

1. FreeBSD – a “classic” kernel

- Used in systems such as Netflix, PlayStation and WhatsApp.
- Considered one of the most reliable codes in its category.

2. The vulnerability

- In the `kgssapi.ko` module, responsible for Kerberos authentication of NFS (RPCSEC_GSS).
- Allows an unauthenticated attacker to trigger a stack overflow during RPCSEC_GSS packet signature verification, opening a path to remote code execution.

3. Claude’s exploit

- Sets up an environment with the vulnerable kernel, NFS and Kerberos.
- Develops multi‑packet shellcode delivery, correctly terminates intercepted kernel threads so the server stays operational.
- Determines stack offsets using De Bruijn sequences, creates a new process via `kproc_create()`, switches it to user mode (`kern_execve()`), clears the `P_KPROC` flag and resets the `DR7` register.

What changed compared to traditional methods

Consequences and lessons

- Attack speed is now comparable to the time it takes cybersecurity specialists to deploy a patch.
- Major OS developers, cloud providers and critical infrastructure must rethink their strategies:
1. Implement AI‑driven security checks as an ongoing process.
2. Monitor intrusion attempts in real time.
3. Quickly move from vulnerability detection to remediation.

What’s next?

Carlini has already applied the Claude‑based vulnerability search scheme to more than 500 critical bugs across various codebases, showing that the key point is not a single exploit but the method of automatic generation and exploitation. This underscores the need to integrate AI into the defense systems of any large organization.