Anthropic’s FlashProve, ChatGPT hits 1.2 billion weekly users, CrowdStrike buys two AI security startups

Compact Conversations for 2026-11-03: 5 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Anthropic releases FlashProve, an open-source prototype for provable inference

Anthropic open-sources FlashProve, a zkSNARK-based prototype that cryptographically proves a response came from a specific model, not one a router or attacker swapped in. For a 400B-parameter MoE model, a single H100 GPU generates proofs at ~110 tokens/second, with ~19 MB proofs per 1,024-token request, roughly 70-95x the GPU time of normal inference.

Why it matters: Enterprises and evaluators using third-party routing services or APIs currently have no way to verify that responses come from the model they paid for, and this prototype demonstrates cryptographic proof is becoming practical.

Source: Anthropic

Number to Know: ChatGPT hits 1.2 billion weekly active users

OpenAI consumer product lead Adam Fry says ChatGPT has hit 1.2 billion weekly active users, adding 200 million in the past two months. Its search volume is now roughly 20% of Google’s, and users create 1.5 billion images per week.

Why it matters: ChatGPT’s rapid growth and search volume indicate it is becoming a primary interface for information retrieval, challenging traditional search engines.

Source: Big Technology

The Feed

CrowdStrike agrees to buy Fiddler AI and Dreadnode for $495 million

CrowdStrike buys Fiddler AI, which helps monitor AI agents in production, for about $400M, and Dreadnode, an automated security AI testing tool, for about $95M, mostly in cash. The deals come as companies deploy AI agents faster than security teams can establish controls.

Why it matters: AI-agent monitoring and security testing are being absorbed into mainstream security platforms, signaling a shift in how enterprises will manage AI risk.

Source: Wall Street Journal

China plans broad AI safety standards covering development through deployment

Bloomberg reports China plans to roll out a broad set of AI safety standards governing development through deployment, targeting domestic AI model makers, following global security incidents.

Why it matters: China’s move could set a regulatory precedent for AI safety that affects global supply chains and model availability.

Source: techmeme.com

Google will cut off search access for students at schools that reject AI

Google says K-12 schools using Google Classroom can now toggle each AI tool individually, but search is integrated with AI, so schools that opt out of AI entirely lose Google search. The move drew criticism from former FTC chair Lina Khan and others.

Why it matters: Google’s dominant position in education technology means this policy could force schools to adopt AI or lose a core service, raising concerns about coercion and student privacy.

Source: AP News

One Thing to Try

Try Kimo, an open-source video editor with no GUI. Everything—cutting, trimming, captions, color grading, export—happens through MCP tool calls, so GitHub Copilot’s agent mode or another MCP client acts as the editor. The tool can render a single frame back to the agent for visual checking, reducing blind iterations. It’s a small Rust binary that runs on modest hardware. The author admits it’s rough, but pointing your coding agent at a headless tool and letting it work end to end is a different workflow worth experimenting with.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Today’s lead is Anthropic releasing FlashProve, an open-source research prototype that tries to answer a question most of us just assume: how do you know an AI response actually came from the model you picked? [thoughtful] Right now, if you call a model through a routing service, nothing stops that router from quietly swapping in a cheaper model, or an attacker from tampering with the token stream. You can’t tell from the output alone.

Host B: FlashProve uses a cryptographic technique called a zkSNARK. In plain terms, the provider publishes a short fingerprint of the model’s weights, then attaches a proof to each response showing those exact weights produced it. The proof reveals nothing about the weights, and anyone can check it quickly on an ordinary CPU without running the model.

Host A: [with emphasis] The performance numbers are the story here. On Llama 4 Maverick, a 400 billion parameter mixture-of-experts model, a single H100 GPU generates proofs at about 110 tokens per second, with proofs around 19 megabytes per thousand-token request. That’s roughly 70 to 95 times the GPU time of running the model normally, which Anthropic calls far less overhead than previously expected.

Host B: [conversational] It’s explicitly a prototype, not a finished product. Contexts are capped at about a thousand tokens, it assumes a mixture-of-experts architecture, and Anthropic says it’s releasing the code, technical report, and formal Lean proofs under Apache 2.0 so others can pick it up and improve it.

Host B: One number to know today: 1.2 billion. That’s ChatGPT’s weekly active user count, according to consumer product lead Adam Fry, speaking at an OpenAI dinner in New York covered by Big Technology. [with a small lift] The company added 200 million weekly users in just the past two months, going from one billion to 1.2 billion in sixty days. Fry also said ChatGPT’s search-like query volume is now roughly one-fifth of Google’s, and that users create 1.5 billion images a week. He noted those figures are OpenAI’s own reporting.

Host A: In the feed, CrowdStrike is buying two AI security startups. The Wall Street Journal reports it agreed to acquire Fiddler AI, which helps monitor AI agents in production, for about 400 million dollars, and Dreadnode, an automated security testing tool for AI, for about 95 million, mostly in cash. [thoughtful] The Journal frames the deals around companies deploying AI agents faster than security teams can establish controls, and CrowdStrike pulling both capabilities into its platform reflects that framing.

Host B: Next up, Bloomberg reports that China plans a broad set of AI safety standards covering everything from development through deployment, aimed at domestic AI model makers. The reporting ties the push to recent global security incidents. Details on timing and enforcement are still thin in what’s been published so far.

Host A: [lighter] And finally, Google is tying search access to AI adoption in schools. Per the AP, K-12 schools using Google Classroom can now toggle each tool individually, including Gemini, but search is integrated with the AI tools, so a school that opts out of AI entirely loses Google search too. That drew quick criticism, including from former FTC chair Lina Khan, and New York City and Los Angeles have temporary bans on generative AI in elementary and middle schools while they write policy.

Host B: One Thing to Try: let an AI agent drive a tool that was built for agents, not humans. A developer shared an open-source video editor called Kimo that has no graphical interface at all. Everything, cutting, trimming, captions, color grading, export, happens through MCP tool calls, so GitHub Copilot’s agent mode or another MCP client acts as the editor. [curious] The trick worth borrowing even if you skip the tool: it can render a single frame back to the agent so it can visually check its own work before moving on, which cuts down blind iterations a lot. It’s a small Rust binary that runs on modest hardware, and the author admits it’s rough, with transitions that pop and fussy audio sync on longer timelines. [conversational] But pointing your coding agent at a headless tool and letting it work end to end is a different workflow, and worth ten minutes of experimentation.

Host A: That’s Compact Conversations for Tuesday. More AI news tomorrow. Until then, happy prompting.