Anthropic announces FlashProve to prove AI answers
Anthropic's research team has announced a mechanism called FlashProve that uses zkSNARKs to prove that an AI's answer was generated by a specific model. The system can produce proofs for about 110 tokens per second per GPU, but it requires 70 to 95 times more GPU time than running the model normally. Anthropic writes that the proof can detect if a relay service swaps a cheaper model or if an…
Key points
- Anthropic's FlashProve can generate proofs for about 110 tokens per second per GPU, but needs 70 to 95 times more GPU time.
- Moonshot AI released Kimi K3.1 Pro with 2.8 trillion parameters and a 1‑million‑token context window, scoring 65.7 on Terminal‑Bench 4.0.
- OpenAI found 948 privilege‑escalation vulnerabilities in Linux, verified them, and proposed fixes in an automated system titled “A Thousand Fewer Paths to Root.”
Moonshot AI released the weights of Kimi K3.1 Pro on Hugging Face, giving users the option to run the model locally. The 2.8 trillion‑parameter model can handle a 1‑million‑token context window and scored 65.7 on the Terminal‑Bench 4.0 programming test, close to Claude Opus 5.5's 66.4. Reddit posts question whether the score is inflated, so the company advises users to test the model themselves before relying on the official report card.
OpenAI announced it discovered 948 privilege‑escalation vulnerabilities in Linux, verified them, and proposed fixes in an automated system titled “A Thousand Fewer Paths to Root.” The report shows AI can accelerate security patching. Other stories include a Hacker News post that Gemini may use paid‑account conversation history for training, and Apple opening a developer waitlist for Siri AI that can read the screen and add custom skills. These developments highlight growing concerns about AI trust, data usage, and new opportunities for voice‑based applications.
Model page: Kimi K3.1 Pro →
The story so far
2 episodes →- Anthropic announces FlashProve to prove AI answersthis story
Is that AI real? Anthropic proves answers with cryptography
note.com · 5 November 2026
Loading the full article…
This text was published by note.com and written by 菅原友和 | AI歴10年目社長. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Bartowski releases Xing4.0-29B-A4B quantized models for llama.cpp · 1 src
- MIT CSAIL: GPT-6 Astra edits Qwen's parameters · 1 src
- PSI proves Aaronson-Ambainis conjecture using Anthropic’s non-public model · 1 src
- Mit develops CheMeleon model that outperforms other methods · 1 src
- OpenAI releases BioBasicsBench benchmark; GPT‑6 Astra scores 66.7% on wet‑lab tasks · 1 src
Comments
via GitHub Discussions