Artificial intelligence safety and research firm Anthropic is working on a new open-source system designed to independently verify which specific AI model produced a given piece of output, according to a company blog post published Tuesday. The system, which researchers have named FlashProve, is built to provide what the team defines as “provable inference”—a cryptographic proof that functions essentially as a mathematical guarantee confirming that a designated output originated from a specific, named model.
The core purpose of the technology is to let end users definitively confirm that the answers, text, code, or data they receive from an AI service genuinely came from the specific model they think they are utilizing, rather than a substitute.
Operating on a principle conceptually similar to DomainKeys Identified Mail, or DKIM, which are the cryptographic email signatures that allow recipients to verify that a message truly originated from its claimed sender, FlashProve is engineered to be robust against potential impostors. This includes defense against malicious actors attempting to disguise one model as another. Crucially, the researchers note that the protection extends even to scenarios where the impostor is the very AI laboratory that built and operates the model in the first place.
In practical terms, the researchers pointed out that a system of this nature could be deployed to prevent an AI company from secretly swapping in a cheaper, less capable model behind the scenes in order to cut operational costs, or conversely, swapping in a more expensive frontier model to artificially inflate its performance benchmarks on evaluations or public leaderboards.
Beyond corporate accountability regarding model substitution, FlashProve could also find utility in verifying third-party claims asserting that a specific, high-stakes task was executed by an advanced, state-of-the-art model. For instance, a software developer might claim that their application’s codebase was thoroughly scanned for security vulnerabilities by a cutting-edge cybersecurity model. Alternatively, a corporation might claim that its overarching business strategy or financial plan was given a rigorous evaluation and greenlighted by an appropriate enterprise-grade model.
In their announcement, the researchers framed these scenarios around a fundamental dilemma facing modern adopters of artificial intelligence: how can a user realistically verify these types of claims when models are typically hosted remotely behind closed API walls?
To solve this challenge, the FlashProve system requires AI labs to generate and publish a unique “fingerprint” for each model version they release to the public. These fingerprints can later be matched directly to specific outputs generated during inference and verified using standard cryptographic methods.
Importantly, the system is designed so that these published fingerprints do not reveal a model’s proprietary weights—the massive arrays of internal numerical parameters that dictate how an AI model behaves, processes information, and generates responses. Commercial AI labs, including Anthropic itself as well as competitors like OpenAI, fiercely guard these weights as their most valuable commercial and intellectual assets. By ensuring that fingerprints do not expose these underlying weights, the system can function effectively even with commercial, closed-source models. Furthermore, the architecture of the system is intentionally lightweight, meaning that the cryptographic verification process can be performed efficiently on a local user device rather than requiring massive cloud infrastructure.
Alongside the initial announcement, Anthropic’s research team released a comprehensive and detailed technical specification document, along with working code for various components of the system. This material has been made publicly available under an open-source license hosted on GitHub, allowing developers, cryptographers, and AI researchers to examine the codebase directly.
Despite the release of working code, the technology is not currently slated for immediate commercial deployment. The researchers emphasized that there are no active plans to integrate FlashProve into Anthropic’s commercial operations or to deploy it across any other major AI lab in the near future.
Instead, the initiative is being presented as an ongoing research effort intended to stimulate broader industry collaboration and technical critique. The researchers explained in their blog post that the publication is meant to share intermediate progress rather than represent a completed research agenda. They noted that they are sharing the work explicitly with the goal that outside developers and researchers will examine what has been built, take the foundational concepts, and improve upon them as the broader AI ecosystem grapples with issues of trust, transparency, and verification.
Comments (0)