AI Autonomously Hacks FreeBSD in Four Hours, Raising Cybersecurity Alarms
- An AI agent using Anthropic's Claude autonomously developed two working kernel exploits against FreeBSD, a widely used and security-hardened operating system, in approximately four hours.
- The exploit targeted CVE-2026-4747, a remote code execution vulnerability in FreeBSD's RPCSEC_GSS module, requiring the AI to independently solve six complex technical challenges without human intervention.
- The same researcher has since identified 500 additional high-severity vulnerabilities using the same AI pipeline, compressing the time and cost required to convert software flaws into operational offensive capabilities.
FreeBSD published a security advisory for CVE-2026-4747, a remote code execution vulnerability in its kernel, crediting 'Nicholas Carlini using Claude, Anthropic' for discovering the flaw. The AI system did not merely flag suspicious code. It developed two working exploits, from scratch, that deliver root shell access on unpatched servers in roughly four hours of compute time.
The vulnerability exists in FreeBSD's RPCSEC_GSS module, which handles Kerberos authentication for NFS servers. A malicious client can trigger a stack buffer overflow without authenticating first.
Exploiting the flaw required the AI agent to solve six distinct problems autonomously: configuring a test environment with the vulnerable kernel, devising a multi-packet strategy to deliver shellcode, cleanly terminating hijacked kernel threads, debugging incorrect stack offsets using De Bruijn patterns, creating a new process from kernel context, and clearing inherited debug registers that caused child processes to crash.
FreeBSD powers Netflix's content delivery network, runs PlayStation's operating system, and forms the infrastructure backbone of WhatsApp. Its 30-year codebase has long been considered a benchmark for mature, audited, and hardened software security.
Industry surveys consistently show the median time to patch critical vulnerabilities in enterprise environments exceeds 60 days. With AI capable of developing working exploits within hours of a vulnerability disclosure, the window between patch availability and exploitation has compressed to near-zero.
The same researcher has since used the Claude-powered pipeline to identify 500 other high-severity vulnerabilities across various codebases. The methodology generalises, and once a capability exists, it diffuses — to nation-state actors, criminal organisations, and independent researchers alike.
Developing a kernel exploit traditionally required weeks of work from highly paid specialists. AI collapses that barrier, with a four-hour compute job now achieving part of what previously required a dedicated team, dramatically shifting the supply curve for offensive capability.
Security teams are urged to integrate AI into their defensive posture through continuous security auditing, real-time monitoring for exploitation attempts, and fundamentally shortening the cycle from vulnerability disclosure to deployed patch.
Regulatory frameworks will also need to adapt. Current compliance regimes assume security is measurable through checklists and periodic audits, and do not account for adversaries whose capabilities compound with each improvement in foundation models.