CVE-2026-72844

The core vulnerability in CVE-2026-72844 is not the hash collision technique used in the proof of concept — that is a delivery optimization, irrelevant to the actual severity. The real issue is that Lean's kernel accepts declarations from metaprograms without verifying that projection targets match their expected structures, and it skips type checking on auxiliary types generated during inductive registration. This creates a genuine inconsistency in the kernel's type theory that passes through the 'maximum kernel checking' path as if it were ordinary user code. This matters because the kernel's type checker is positioned as the final arbiter of consistency, yet metaprogram-driven declaration paths corrupt the type theory before that checking occurs. When the kernel admits an inconsistent inductive declaration, the resulting environment allows derivation of False — and from False, any proposition follows. There is no crash, no error log, no runtime signal. The corruption is silent and mathematical. The severity of this vulnerability hinges on a single architectural question that the CVE description does not answer: can the inconsistent proof context survive serialization into a .olean file and be instantiated in a downstream Lean environment? If .olean export preserves the inconsistent state, then the attack surface is supply chain — a malicious dependency publishes a poisoned package, and every downstream consumer imports the corruption silently. If the inconsistency only exists within the originating process, the vulnerability is real but contained to in-process metaprogram execution. The practical implication is that 'maximum kernel checking' performs a guarantee it cannot deliver when metaprograms bypass the same structural validation required of user-level code. The setting implies a trust decision — that declarations from any source are verified — but the implementation violates that contract. Users who enable this setting are making a trust decision they cannot enforce. The fix is not merely patching the projection and inductive checks; it requires formalizing which code paths produce declarations the kernel must accept, and whether metaprogram-initiated additions should be privileged or go through the same validation as user declarations. Until that specification exists, every .olean artifact becomes a potential carrier of silent logical corruption.

Reviewed through automated stages and approved by a human before publication.