Radar / AI security / Anthropic FlashProve
Anthropic FlashProve: cryptographic proof of which AI model actually answered
Anthropic open-sourced FlashProve, a research prototype that uses zkSNARKs to prove a response came from one specific committed model without revealing its weights, demonstrated on Llama 4 Maverick at about 110 tokens per second of proof generation and released under Apache 2.0.
Why it matters
The first practical zero-knowledge proof system for model identity at 400B scale: it lets users verify a provider ran the model it claims, with no trusted hardware and no weight disclosure.
What you could build with it
An enterprise team buying hosted model access could require providers to publish FlashProve-style weight commitments and attach proofs to billed responses, giving auditors a checkable receipt that the premium model was actually used.
Does it hold up?
Too early to judge: Anthropic calls it an unmaintained research artifact with no deployment plans, 70-95x proving overhead, and a 1,024-token cap, so it is a verification primitive to study, not a production tool yet.
Built with Anthropic FlashProve
- FlashProve puts cryptographic receipts behind hosted AIartiverse · A detailed walkthrough of the commitment-and-proof design, the 18.78 MB proofs, and the CPU-only verification step.
- Anthropic releases FlashProve to verify AI model identity using cryptographic proofstheoutpost · News report on the launch, the Llama 4 Maverick demonstration, and Anthropic's note that it has no deployment plans yet.
- Anthropic releases FlashProve to verify AI model responsestronweekly · Coverage of the 70-95x proving overhead, the 1,024-token context cap, and the Apache 2.0 release of code and formal proofs.
Learn more
First spotted on article: source.
More AI security
NVIDIA SkillSpector: security scanner for AI agent skills, trending againNVIDIA's open-source SkillSpector scanner is trending on GitHub today: it audits AI agent skills before installation…security · JEV 0.75Shannon: open-source AI pentester that reports only vulnerabilities it can actually exploitCLI agent that combines source-code analysis with live exploitation against web apps and APIs under a 'no exploit, no…security · JEV 0.71AiSOC: open-source AI security operations center with agent triageSelf-hosted SOC that fuses alerts into incidents and hands them to LLM agents that record every prompt and tool call…security · JEV 0.71iFixAi: independent auditing of AI agents in under 120 secondsiFixAi audits whether an AI agent is doing what it is supposed to do, run by a human or by the agent itself, answering…security · JEV 0.68BLA: adversarial defense shield that poisons, exhausts and hallucinates AI reverse-engineering agentsTypeScript shield protecting binaries, Electron apps, JS bundles, Python and WebAssembly against AI…security · JEV 0.47Anthropic sandbox-runtime: OS-level sandboxing for AI agents, no containers requiredAnthropic's sandbox-runtime (srt) is trending with developers this week: an open-source tool that wraps any process in…security · JEV 0.45
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.