Anthropic’s Claude Fable 5 generated a complete, bootable NT-compatible Windows kernel written in Rust called ntoskrnl-rs from an empty directory in just 38 minutes of active model work, raising profound questions about AI-authored trust and the future of critical infrastructure security.

Documented by security researcher Matt Suiche and Tolmo’s threat research on June 22, 2026, the project set out to rewrite ntoskrnlThe Windows NT kernel, in Rust.

Fable 5 handled the core scaffolding in a single contiguous session, producing approximately 5,100 lines of code across 27 files spanning the scheduler, memory manager, trap and interrupt machinery, object manager, and I/O manager.

The kernel successfully booted in the QEMU emulator and passed all 14 in-kernel self-tests, exiting with the project’s standing pass contract: exit code 33.

Claude Fable 5 Wrote Windows Kernel Code

The wall-clock time across the full session was roughly four and a half hours, but most of that was the human operator away from the keyboard; the actual model-active work took 38 minutes.

What distinguishes this from simple code generation is Fable 5’s demonstrated capacity for unsupervised systems reasoning. The model caught two critical low-level bugs mid-generation without human intervention:

• EOI ordering bug: It identified that the end-of-interrupt signal must be issued before a potential context switch, as preemption mid-dispatch would deadlock the local interrupt controller.
• IRQL emulation bug: When host tests returned 11/12, Fable diagnosed that the interrupt request level (IRQL) emulation used a single global atomic across test threads, corrected it to a per-thread thread_local variable mirroring real per-CPU behavior, and passed 12/12.

The model also left architectural commentary embedded in the code explaining why the NT GDT selector ordering matches the IA32_STAR MSR format demonstrating forward-looking ABI reasoning, not pattern matching, reads Tolmo’s report.

Fable 5 authored roughly 40% of the project’s from-scratch code in only 3% of the total turns. The remaining 97% of turns eight days of iterative, debug-heavy bring-up ran on Claude Opus 4.8, which expanded the kernel to load unmodified Windows kernel drivers and run real Windows binaries including sort.exe, choice.exe, and cmd.exe.

The model split was deliberate. Fable 5 carries aggressive cybersecurity safety classifiers broad enough to trip on adjacent defensive work. Notably, Fable shipped on June 10, 2026, as the public version of Anthropic’s Mythos cybersecurity model, and within days, a US government export-control directive forced Anthropic to suspend access entirely.

The kernel boots. It is not yet trusted. Fable 5 itself named the gap unprompted, flagging the dispatcher lock hand-off, spinlocks, and DPC queue as the highest-risk paths, and recommending loom for exhaustive concurrency exploration and Miri for undefined behavior detection.

This is the critical security implication: authoring capability has outpaced verification. A model can produce the Trusted Computing Base of an x86_64 kernel faster than any human team can audit it.

Until tooling like formal verification, property testing, and concurrency model checkers can close that gap, an AI-authored kernel remains a booting artifact of unknown correctness and unknown correctness has no place in a TCB.

The internet’s critical infrastructure runs on aging C codebases maintained largely because rewriting a TCB has historically been too costly and too risky. An AI-authored Rust kernel represents a double lever: Rust eliminates the memory-safety bug classes that dominate OS CVEs, while an AI model eliminates the human-cost bottleneck of the rewrite itself.

Once verification tooling matures, the economic case for leaving legacy C in place collapses and large portions of the stack become candidates for AI-driven, memory-safe rewrites.

Follow us on Google News, LinkedIn, and X to Get More Instant Updates.