Tag
Anthropic announced FlashProve, a provable inference system using zkSNARKs that lets users cryptographically verify they are being served the exact AI model and weights promised, without revealing model weights. Built with Claude's help, it is claimed to be an order of magnitude faster than existing open-source solutions and has implications for model integrity, auditing, and future AI regulation.
The paper introduces ExecCert, a release-time certification layer that bridges the gap between mathematical certified machine-unlearning guarantees and the concrete finite-precision artifacts actually deployed, including an incremental Retraining-Reference Release Verification for sequential deletion requests.
A technical blog post presenting a reproducible pipeline to fingerprint whether an LLM was trained from scratch or derived from existing open weights, using architecture, tokenizer, and weight comparisons.
Moonshot AI has open-sourced the Kimi Vendor Verifier (KVV), a tool designed to help users verify the accuracy and correctness of inference provider implementations for open-source models like Kimi K2. It uses six critical benchmarks to detect infrastructure-level issues such as KV cache bugs, quantization degradation, and parameter misuse.