Anthropic — Claude discovers FreeBSD kernel RCE, writes full exploit

AI relevance: Claude demonstrates autonomous vulnerability discovery and exploit development capabilities, fundamentally changing the economics of offensive security research and raising questions about AI agent governance.

Nicholas Carlini from Anthropic has demonstrated a groundbreaking capability: using Claude to both discover a critical FreeBSD kernel vulnerability and develop a complete remote code execution exploit.

What happened

  • CVE-2026-33017: Integer overflow in FreeBSD's RPCsec GSS implementation (kgssapi.ko)
  • Attack surface: NFS servers with RPCsec GSS enabled (port 2049/TCP)
  • Impact: Remote kernel code execution leading to full root compromise
  • Credits: "Nicholas Carlini using Claude, Anthropic" per FreeBSD security advisory

Why this matters

  • Autonomous discovery: Claude found the vulnerability through source code analysis
  • End-to-end exploitation: Developed complete ROP chain and payload delivery
  • No KASLR bypass needed: FreeBSD 14.x lacks kernel address space layout randomization
  • Economic shift: Dramatically reduces cost of vulnerability research

Technical details

  • Vulnerability: Integer overflow in array index calculation
  • Exploit chain: Crafted RPC packets → kernel overflow → ROP → root shell
  • Payload: Adds public key to authorized_keys for persistent access
  • Reliability: Works reliably against default FreeBSD 14.x configurations

Broader implications

  • Defender advantage? AI can find vulnerabilities before attackers
  • Scale problem: Potential flood of CVEs overwhelming maintainers
  • Governance gap: No framework for responsible AI vulnerability research
  • Arms race: Attackers will use similar AI capabilities

What to do

  • Patch immediately: FreeBSD-SA-26:08.rpcsec_gss advisory released
  • Network segmentation: Isolate NFS servers from untrusted networks
  • Monitor activity: Watch for unusual RPC traffic patterns
  • Review AI governance: Establish policies for AI-assisted security research

Sources

This represents a fundamental shift in cybersecurity capabilities — AI agents can now autonomously discover and weaponize vulnerabilities at scale.