Anthropic Develops FlashProve to Expose Hidden Claude Usage
Artificial intelligence developer Anthropic is building an open-source tool that could expose companies secretly using its Claude models to generate responses while falsely claiming the output came from their own technology. The system, known as FlashProve, relies on cryptographic verification to prove which specific artificial intelligence model produced a given text response. Although the company has not explicitly linked the initiative to foreign competitors, the software addresses practices that Anthropic has previously attributed to Chinese laboratories.
The development of FlashProve follows a series of public accusations by Anthropic regarding unauthorized access to its technology. In September 2026, the company stated that Moonshot AI had secretly routed nearly 300,000 customer requests through Claude over a ten-day period. These developments highlight ongoing disputes within the artificial intelligence sector regarding intellectual property, API misuse, and the unauthorized extraction of proprietary model capabilities.
Chinese AI Companies Accused of Unauthorized API Routing and Distillation
The friction between Anthropic and various Asian artificial intelligence developers intensified throughout 2026 following detailed public accusations. According to Anthropic, Moonshot AI used approximately 5,380 fraudulent accounts, which largely appeared to operate from Singapore and Japan, to access Claude and present the outputs as Kimi customer responses. Anthropic also stated that Moonshot saved conversations and extracted Claude reasoning data to improve its own proprietary models.
Earlier in February, Anthropic accused DeepSeek, Moonshot, and MiniMax of conducting large-scale model-distillation campaigns. These operations allegedly involved more than 16 million exchanges conducted through roughly 24,000 fraudulent accounts. Model distillation involves using responses generated by a more capable artificial intelligence system to train a competing model. While the technique has legitimate research applications, Anthropic argues that unauthorized extraction violates its terms of service.
In September, Anthropic also accused Alibaba, Zhipu, and Xiaomi of participating in similar activities, alongside allegations that DeepSeek had secretly forwarded customer requests to Claude. These assertions remain corporate accusations rather than independently established findings. However, they demonstrate the growing need for technical verification methods within the global technology sector.
How Anthropic's FlashProve Cryptographic System Works
Introduced on November 2, 2026, FlashProve functions as a system for provable inference, enabling users to mathematically verify the origin of an artificial intelligence response. Ordinarily, consumers and enterprise clients must rely entirely on the statements of service providers when claiming a specific model generated an answer. Without technical verification, providers can engage in deceptive practices to reduce operational expenses or improve perceived performance.
For example, an artificial intelligence provider might secretly substitute an expensive, highly capable model with a cheaper alternative to save money. Alternatively, a company could deploy a more powerful system during formal benchmark testing to make its commercial products appear superior to those of competitors. FlashProve aims to eliminate this misrepresentation through cryptographic proof mechanisms.
Under the proposed framework, artificial intelligence providers would publish a unique cryptographic fingerprint associated with each distinct model. When a model generates a response, the provider produces a mathematical proof linking that specific output to the published fingerprint. Users can subsequently verify whether the response originated from the advertised model. If a different system produced the text, the verification process fails.
Verifying Models Without Revealing Proprietary Weights
A central design challenge in creating verification tools for proprietary artificial intelligence involves protecting intellectual property. Companies such as Anthropic and OpenAI keep the internal weights of their leading models strictly confidential. Model weights represent the complex numerical values that determine how an artificial intelligence system processes information and formulates responses.
FlashProve resolves this tension by allowing providers to publish cryptographic fingerprints without exposing their proprietary model weights. The system is designed to support lightweight verification on standard local devices, though generating the initial cryptographic proofs demands substantial computing power. Participation requires artificial intelligence providers to willingly publish fingerprints and generate the accompanying mathematical proofs for their services.
Detecting Misleading Claims in Professional and Benchmark Tasks
Beyond identifying which model generated a specific response, FlashProve could help verify claims made by businesses using artificial intelligence for professional tasks. For instance, a software developer might claim that an advanced cybersecurity model scanned its codebase for vulnerabilities. Similarly, a corporate enterprise might advertise that a powerful artificial intelligence system reviewed a complex business plan.
FlashProve could provide mathematical evidence confirming which model actually performed the task. The technology could also assist researchers in verifying the exact models used during benchmark evaluations, preventing companies from secretly substituting superior systems to skew results. Nevertheless, proving a model identity does not guarantee that its answers or conclusions are factually accurate.
Technical Testing and Hardware Requirements for FlashProve
Anthropic has demonstrated the technical viability of FlashProve by testing the system on Meta's Llama 4 Maverick, a large language model containing 400 billion total parameters. Operating with four-bit quantized weights and a single Nvidia H100 graphics processing unit, the system generated proofs at a rate of approximately 110 tokens per second.
For a request containing 1,024 tokens, the generated proof measured approximately 19MB and could be verified locally on a central processing unit within seconds. However, proof generation required roughly 70 to 95 times more graphics processing unit computation than standard inference using vLLM software. This high computational overhead remains a significant barrier to widespread commercial adoption.
Open-Source Availability and Current Limitations
Anthropic has released FlashProve's technical specification, mathematical research, and working code on GitHub under an open-source license. The company acknowledges that the project remains experimental and subject to limitations regarding model computations, supported output types, and the formal relationship between its optimized implementation and the underlying protocol.
Anthropic has not announced any immediate plans to integrate FlashProve directly into Claude or its other commercial offerings. Instead, the company has made the technology available for independent researchers to study, test, and refine. While the tool cannot independently identify unauthorized model distillation or track training datasets, it provides a foundation for greater accountability regarding model deployment claims.
