Understand this story
Anthropic has released FlashProve, an open-source research prototype that uses cryptographic proofs to verify that a given output actually came from a specific model. According to a TechCrunch AI report, the system requires AI labs to publish a fingerprint for each model that does not expose its weights, which can prevent vendors from secretly swapping in a cheaper or more expensive model, and can also verify third-party claims that a task was completed by a certain model; by design it is lightweight enough that verification can be done on a local device. Another report says that on Meta's Llama 4 Maverick (400B parameters), a single Nvidia H100 generates proofs at 110 tokens/second, each about 19MB, and an ordinary CPU takes about 17 seconds to verify one; previously the only known system capable of proving that model, VerInf, generated 93.6GB proofs and verification took nearly a day. The system was built by Claude, and because the code is difficult to audit, its security guarantees are proven mathematically in Lean 4. The researchers have released a technical specification and some open-source code; the research is still ongoing, and there are currently no plans to deploy it at Anthropic or other labs.
AI-generated from public reports
Latest development · Report
Anthropic develops FlashProve system to cryptographically prove AI model identityDevelopments
1 confirmed developments
Anthropic releases open-source research prototype FlashProve
Source:TechCrunch AI
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 4
TechCrunch AI
Anthropic develops FlashProve system to cryptographically prove AI model identity
Anthropic is developing an open-source system called FlashProve that uses cryptographic proofs to verify that a given output actually came from a specific model. The system requires AI labs to publish a fingerprint for each model that does not expose its weights, which can prevent vendors from secretly swapping in a cheaper or more expensive model, and can also verify third-party claims that a task was completed by a certain model; by design it is lightweight enough to allow verification to be completed on local devices. The researchers have published the technical specification and part of the open-source code, and the research is still ongoing, with no plans for deployment at Anthropic or other labs at this time.
Rohan Paul
Anthropic open-sources FlashProve, which can verify on an ordinary CPU which model a response came from
Anthropic has open-sourced FlashProve, enabling servers to mathematically prove that a given response came from a specified model, with verification completable on an ordinary CPU. On Meta's Llama 4 Maverick (400B parameters), a single Nvidia H100 generates proofs at 110 tokens/second, each about 19MB, and an ordinary CPU verifies one in about 17 seconds; the only previously known system capable of proving this model, VerInf, generated 93.6GB proofs, with verification taking nearly a day. The system was built by Claude, and because the code is difficult to audit, its security guarantees are mathematically proven in Lean 4.
Anthropic
Anthropic releases FlashProve, which can verify the source of model responses and open-sources the code
Anthropic has announced FlashProve, a new research direction in verifiable inference. Built on existing cryptographic techniques, it lets developers verify whether a given response came from a specified AI model, and Anthropic says it is significantly more efficient than previous open-source methods. The proof can be used to make it harder for attackers to hijack conversations, and it also allows safety evaluators to confirm that the model being tested is the same as the model serving the public. Anthropic has open-sourced the FlashProve code and lists collaborators including Alessandro Chiesa.
Nov 3
Unite.AI
Anthropic Releases FlashProve, Using Cryptography to Prove the Provenance of Model Reasoning
On November 2, 2026, Anthropic released FlashProve, an open-source research prototype that uses zkSNARKs to prove that a given response came from committed model weights without revealing the weights themselves; for Meta's Llama 4 Maverick, it reports a proof generation speed of 110 tokens/s. A single H100 plus a 24-core CPU can prove 24 requests in about 223 seconds, generating an 18.78 MB proof per request, with the offline commitment phase taking about 39 minutes; after downloading a commitment of about 6.5 MB, the client verifies on CPU in 17 seconds with 24 threads and 21 seconds single-threaded. Compared with FP8 inference running vLLM on 8 H100s, the single-GPU time required for proving is about 72 times higher, and about 94 times higher under maximum load. The code is released under Apache 2.0, and version 0.16.0 is labeled as an unmaintained research artifact with no independent security audit yet; proofs cover only the integer-weight version under greedy decoding, with a context limit of 1024 tokens.
Heat trend for this story
- Current
- 33
- Period peak
- 34 Nov 4 07:00
- 24-hour change
- —
Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.