Anthropic FlashProve Verifies AI Model Identities

Anthropic published a blog post Tuesday describing a tool meant to answer a question that's gotten harder to settle as AI products multiply: did the output you got actually come from the model you were told you were using. The system is called FlashProve, and it's built around what the researchers term "provable inference," a cryptographic proof that ties a specific output to a specific model, the same way a signed email ties a message to its sender.

The stakes aren't abstract. Anthropic's researchers point to a scenario where an AI company quietly swaps in a cheaper model to save money, or the reverse, slots in a pricier one to make benchmark results look better than what customers would normally get. FlashProve is designed to make that kind of substitution detectable, closing off a gap that's been largely unpoliced since AI labs started selling access to models as a product rather than shipping them as fixed software.

The approach also extends to third-party claims, like a vendor saying a piece of code was checked by a specific security-focused model, or a company claiming a particular system signed off on its business plan. Making those claims checkable requires AI labs to publish something new alongside each model release.