Claude wrote a full FreeBSD remote kernel RCE with root shell

Role of LLMs in finding vs exploiting bugs

• Commenters stress the exploit was written from an existing advisory, but the underlying bug itself was also originally found with help from an LLM, according to the FreeBSD security notice.
• Several see this as moving the line on what “only humans can do,” especially around kernel exploit development and ROP-style work.
• Others emphasize the human-in-the-loop nature: the shared prompt history shows lots of steering, nudging, and iteration rather than a single-shot “write full exploit” request.

Effectiveness and evidence of AI bug-finding

• Some claim LLMs are already “expert level” at finding vulnerabilities, citing talks, AI CTFs, and tools like Xbow.
• Examples mentioned: internal use at companies discovering dozens of CVEs, framework maintainers finding several, and browser vendors reportedly finding hundreds of issues with LLM help.
• Skeptics push back, asking for harder evidence (e.g., bug bounty stats) and noting that some public red-team writeups looked underwhelming.
• There’s disagreement about how to interpret earlier Anthropic red-team work; one side calls it basically a glorified search for unsafe functions, others point to concrete, acknowledged browser CVEs as counterevidence.

Offense vs defense and the CVE flood

• Many see cheaper, automated CVE discovery as a net positive: defenders and maintainers can find bugs that previously only well-funded attackers would.
• Others worry about a flood of low-impact or duplicate CVEs, and about fixing becoming the bottleneck since patches are often non-trivial and style-sensitive.
• There’s debate over whether LLMs will help more on defense (finding and fixing) or primarily empower attackers, turning this into an arms race.

Fuzzing, testing, and workflows

• Commenters describe using LLMs to:
  • Design fuzzing strategies and harnesses.
  • Analyze crash logs and iterate on tests.
  • Reverse engineer binaries/firmware with tools like Ghidra, radare2, and dynamic instrumentation.
• Several advocate simulation-style, standalone tests with rich logs, specifically to feed AI systems that can generate remediation guidance.

Exploit scope and FreeBSD security posture

• The exploit’s attack surface is relatively narrow: a specific NFS+Kerberos setup; one exploitation path further assumes SSH access for key injection.
• Discussion notes that FreeBSD lacks some mitigations (e.g., KASLR and certain stack protections in this context), though there’s confusion about existing ASLR knobs vs true kernel ASLR.
• Some argue KASLR is limited but still part of a “defense-in-depth” onion.

Meta: hype, newsworthiness, and risk

• Some find it concerning or overhyped that this is still “news”; they see it as an expected capability of frontier LLMs.
• Others see the autonomy angle—agents chaining bugs into working exploits—as what truly worries enterprises and drives calls for governance and safety.
• A few view the thread as unpaid marketing for AI vendors; others reply that the capabilities are now visible enough to “believe your eyes.”