179 points83 comments
Anthropic adds cryptographic proof to LLM responses
What is Progress on provable inference for large language models?
Anthropic released a system that generates zero-knowledge SNARK proofs showing each answer was produced by a specific model. The proof can be verified without revealing the model’s weights or the prompt.
Who it is for
Security-focused developers can use it to guarantee model integrity and prevent tampering in AI services.
How to try it
Read the research blog at https://www.anthropic.com/research/provable-inference