AI Weekly Malaysia

Anthropic says it is building FlashProve, an open-source system for "provable inference" — a cryptographic proof that a given output came from a specific model — described in a blog post published Tuesday. Labs would publish a per-model "fingerprint" that can be matched against outputs without revealing model weights, which is what makes it usable for closed commercial models. The stated use cases are catching a lab secretly swapping in a cheaper or pricier model, and verifying third-party claims like "a frontier cybersecurity model scanned this library." No release date, benchmark, or proof sizes were given in the available text.

Why: If fingerprints ship, "which model actually answered this request" becomes something you can check rather than trust — relevant if you route prompts through resellers, aggregators, or white-label APIs and bill clients on the claim that a frontier model did the work. Until a spec and reference implementation exist, this is a stated direction, not something you can build against; don't write it into contracts or compliance docs yet.