Anthropic is working on an open source system to verify which AI model produced a given output, according to a blog post published Tuesday. The system, called FlashProve, provides what researchers call “provable inference” — a cryptographic proof that a given output came from a specific model, allowing users to confirm the source of the answer they receive.
Similar to cryptographic email signatures, FlashProve can protect against impostors trying to disguise one model as another, even when the impersonation originates from the model’s own lab. This capability could prevent an AI company from secretly swapping a cheaper model to cut costs or substituting a more powerful model to inflate performance results.
The system also enables third‑party verification of claims that a specific task was performed by an advanced model. For example, a developer might assert that a software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that a business plan was approved by an appropriate model. FlashProve offers a way for users to validate such assertions.
To function, AI labs must publish a “fingerprint” for each released model. This fingerprint can later be matched to outputs and verified cryptographically without exposing the model’s internal weights—the proprietary parameters that define its behavior. By keeping the fingerprints lightweight, verification can be performed on a local device, making the system practical for commercial deployments.
As part of the announcement, the researchers released a detailed technical specification and working code for various components of FlashProve under an open source license on GitHub. The project is still in active research, and there are no plans to deploy it within Anthropic or any other lab at this time.
The team emphasizes that this post shares intermediate progress rather than a completed agenda, and they invite others to examine, adopt, and improve the work.
