Anthropic Releases FlashProve Code as Open Source — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic said in a Tuesday blog post that it is developing the open-source FlashProve system, whose “provable inference” mechanism provides cryptographic proof that an output came from a specified AI model.
- FlashProve is designed to resist model substitution by the lab that built and runs the model, including secret swaps to a cheaper model to cut costs or a more expensive model to inflate performance results.
- Anthropic and OpenAI keep model weights private, while FlashProve’s fingerprints are designed to verify outputs without revealing those weights, allowing closed commercial labs to use the system and local devices to perform checks.
- FlashProve is designed to verify third-party claims, including a developer’s claim that a frontier cybersecurity model scanned a software library or a company’s claim that an appropriate model approved a business plan.
- GitHub hosts a detailed technical specification and working code for FlashProve under an open-source license, with the code covering various elements of the system.
- Anthropic says research is ongoing, no lab plans to deploy FlashProve, and the release is intended for others to take and improve rather than serve as a completed research agenda.
Why it matters: FlashProve’s design calls for one cryptographic fingerprint per model release, letting developers and companies check model substitutions and third-party claims without exposing proprietary weights. Anthropic states that no lab plans deployment, so the immediate material outcome is an open research codebase for outside researchers—not a new commercial verification service.
Ask SkimNews
