Anthropic Open-Sources FlashProve for Cryptographic Model Verification
Anthropic is developing FlashProve, an open-source system designed to verify which AI model generated a specific output using cryptographic proofs. By employing model fingerprints that do not reveal weights, the system enables lightweight local verification to prevent providers from secretly swapping models or faking benchmarks. The project is currently in the research phase with code and specifications publicly available.
Related Entities
Full Report
IT Home reported on November 4 that Anthropic is developing an open-source system designed to verify which AI model generated a specific output. According to a blog post published by the company on Tuesday, the system, named FlashProve, aims to achieve what researchers call "provable inference": providing mathematical guarantees via cryptographic proofs that a particular output indeed originated from a specified model. This allows users to confirm whether the answers they receive were truly generated by the model they believe they are using.

As understood by IT Home, FlashProve works similarly to encrypted email signatures. While the latter helps recipients confirm that an email actually came from its claimed sender, FlashProve defends against attempts to disguise one model as another—even if the entity attempting the disguise is the very AI company that developed and operates the model, bypassing verification remains difficult.
Researchers specifically noted that this system could prevent AI companies from secretly switching to cheaper models to cut costs, while also avoiding scenarios where companies covertly switch to more expensive models to artificially inflate performance benchmarks.
FlashProve can also be used to validate third-party claims that a task was completed by an advanced model. The researchers wrote: "Developers might claim their software library has been scanned for vulnerabilities by a cutting-edge cybersecurity model; a business might claim its commercial plan has been approved by an appropriate model. How can users verify these assertions?"
To achieve this, AI labs need to provide a "fingerprint" for each released model. These fingerprints can then be matched with specific outputs and cryptographically verified to confirm their origin.
Crucially, these fingerprints do not leak model weights—the internal parameters that determine model behavior. Closed-source model developers like Anthropic and OpenAI typically protect these parameters as trade secrets. Therefore, AI labs can publish model fingerprints without relinquishing their most valuable assets, meaning FlashProve is equally applicable to commercial closed-source models.
Additionally, the system is designed to be lightweight, allowing users to perform verification directly on local devices.
As part of this release, Anthropic researchers published detailed technical specifications along with runnable code for various components of the system. This code has been made available on GitHub under an open-source license.
Currently, research related to FlashProve is ongoing, and Anthropic has no immediate plans to deploy the system within its own or other AI labs.
In the announcement, the researchers explained: "This research article aims to share interim progress rather than present a completed research program. We are publishing these results with the explicit goal of inviting others to scrutinize our work, build upon it, and further improve it."