Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

Anthropic's FlashProve: a cryptographic proof that a specific model produced a specific answer

Curated by Alfredo Borrás · Plymouth AI Lab

Anthropic publishes an open-source system that proves which model produced an answer, Postman's desktop app is caught installing plugins into coding agents, Revolut says self-hosting open models cut its coding bill by about 90%, and critics take a closer look at yesterday's weight-editing paper.

The story today

Anthropic's FlashProve: a cryptographic proof that a specific model produced a specific answer

On 2 November Anthropic's Frontier Red Team, working with researchers at EPFL, MIT and others, published FlashProve, an open-source prototype for what it calls provable inference. The provider first publishes a short commitment, a fingerprint of the model's weights that does not reveal them. Each reply can then come with a proof that these exact weights, run on this prompt, produce exactly these tokens. A checker needs no GPU and never sees the weights.

The test model was Meta's Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model in 4-bit form, not a Claude model. On one H100 GPU the prover generated proofs at about 110 tokens per second, with each proof about 19 MB for a 1,024-token request. Checking a proof took 17 seconds on 24 CPU threads. Anthropic measured proving at roughly 70 to 95 times the GPU time of serving the same requests with vLLM, a common open-source inference engine, so this is not a cheap add-on.

The authors list limits themselves. The proof covers an integer version of the model under greedy decoding, so a floating-point or sampled reply may not match exactly. It handles contexts of up to 1,024 tokens, assumes a mixture-of-experts design, and hides the weights but not the model's architecture. Anthropic calls the post intermediate progress and releases the code, Lean formal proofs and a technical report under Apache 2.0.

Why it matters: If you buy model access through a reseller or router, there is now a worked example of how a provider could prove you got the model you paid for. It is a research prototype with a heavy cost, so expect it first in audits and evaluations, not in everyday serving.

Sources: Anthropic: Progress on provable inference for large language models

Highlights

  1. Postman's desktop app was reported installing plugins into Claude Code and Codex, with a startup hook that steers agents to Postman

    A GitHub issue opened on 25 October shows Postman's own log recording an automatic install of its Claude Code and Codex plugins, plus a command-line tool, after an app update on version 12.33.4. The Codex plugin adds a session-start hook that tells the agent to load Postman's skill for any non-trivial API engineering task, and the installer code includes a step that approves that hook. The reporter says nothing shows data was uploaded. On 3 November a Postman representative replied that the bundle was meant to ease setup, said the app can uninstall each part, and promised a clearer opt-in and a review of the hook approval and startup behaviour.

    toolssecuritySource: GitHub: postmanlabs/postman-app-support issue 14370

  2. Revolut says moving 2,000 engineers to self-hosted GLM cut its AI coding inference costs by about 90%

    The Register reports a talk at the AI Engineer conference in which Revolut's David Abitbol said the company serves GLM-5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers. He put self-hosted cost at about $0.30 per million tokens, around a tenth of Codex or Claude Code, and cited 60% better coding time and 70% better lead time. The largest gains, he said, came from agents debugging and handling Jira and Confluence tickets once tokens stopped being rationed. All figures are the company's own, the article gives no side-by-side quality comparison, and image tasks still go through a router to Gemini or GPT Luna because GLM is text only.

    industryopen-weightsSource: The Register

  3. spaCy's author says the Astra weight-editing post blends several different methods, and only three experiments read raw weights alone

    Following yesterday's issue on GPT-6 Astra editing Qwen3.5-0.8B, Matthew Honnibal tabulated every experiment in the original post. The 'just by looking at the raw weights' setting, with no tools and no running the network, appears only in experiments 1, 2 and 4, on networks of 17K to 172K weights. Every experiment on a real language model, including the Qwen maths result that went from 7.9% to 48.1% on MATH-500, used Python, and the Qwen run also executed the model on up to 50 batches of practice problems. He concludes the Qwen result shows Astra can devise and carry out tasks a human machine-learning engineer could do, not that it reads weights alone. This is a critic's blog post and he did not rerun the experiments.

    researchinterpretabilitySource: Matthew Honnibal, honnibal.dev

Concept of the day

Zero-knowledge proof

A zero-knowledge proof lets one party convince another that a statement is true without revealing anything beyond the statement itself. In FlashProve the statement is that a committed model, run on a given prompt, outputs a given list of tokens. The provider publishes the commitment to its weights once and attaches a proof to each response, and anyone holding the commitment can check the proof without seeing the weights. A reply from a different model cannot produce a proof that passes against that commitment.

Source: Anthropic: Progress on provable inference for large language models

Models released

ModelMakerSizeNotes
P-H3-ImagePrunaAI (based on MiniMax-H3)33.1B-parameter transformer plus a 4B text encoderText-to-image and editing with up to 8 input images, 8 steps with no classifier-free guidance, 8-megapixel output. Needs about 101 GB of GPU memory in bf16 with the prompt upsampler, and a separate ComfyUI version runs in 21 GB. The MiniMax H3 Community License excludes the EU, UK, South Korea and US without a separate licence.

Local models

  1. Simon Willison runs Qwen 3.8 27B on a laptop to identify games from raw network weights, in 46 to 65 minutes per try

    Willison fed the no-tools prompts from the Astra paper's repository to Qwen 3.8 27B, a 17 GB model in LM Studio. It named Pac-Man after 46 minutes 12 seconds and 43,506 reasoning tokens at 72% stated confidence, Pong in 1 hour 4 minutes 54 seconds, and Breakout in 52 minutes 25 seconds. It failed the Freeway prompt, calling it a Pac-Man-like game. He treats it as a fun result on tiny networks and says the wider commentary goes too far.

    localinterpretabilitySource: Simon Willison's Weblog

Research

  1. Teaching Qwen Why: Anthropic's ethical-advice training cuts agentic misalignment from 61% to 6% on Qwen 2.5 32B

    Second Look Research fellows partly replicated Anthropic's 'Teaching Claude Why' on Qwen 2.5 base models. Fine-tuning on transcripts of an assistant advising users through ethical dilemmas cut the average misalignment rate in an email-agent test from 61% to 6% on the 32B model and from 69% to 6% on the 72B model. Fine-tuning on fictional stories of well-behaved AIs helped less, 43% on 32B and 63% on 72B, and the model that wrote the training data mattered more than any other choice. It covers a subset of Anthropic's experiments on older open models and the authors say it does not explain why the methods work.

    researchalignmentSource: LessWrong: Teaching Qwen Why

  2. An EEG study of 11 developers finds that coding with an agent left them less invested without lowering mental load

    Oregon State researchers had 11 participants each use GitHub Copilot Chat and Cursor Agents in 20-minute sessions, once on their own project and once on a legacy codebase, while recording EEG. With the agent, a measure of motivation towards the task stayed near neutral even on their own projects, and a cognitive-load measure was higher than with Copilot Chat. They also moved from reading code to checking the running app, which they did in 83% of work cycles against 47%. The authors call this the vibe paradox. It is a small exploratory study of students at one university, and EEG measures of load and engagement are indirect.

    researchhumansSource: arXiv: The Vibe Paradox

Also today

  1. China's planner says it will speed up a single AI law and asks developers to report incidents

    ForkLog, relaying a 3 November briefing by the National Development and Reform Commission and Bloomberg's coverage, reports that an official said China will accelerate a 'basic comprehensive' law covering AI from development to deployment. Developers and operators are expected to disclose security threats and report incidents and vulnerabilities to the authorities, and the rules will include a transition period of unstated length. No draft text or adoption timeline has been announced, and the report does not say the incident request carries legal force.

    policychinaSource: ForkLog

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI automation services for logistics

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call