- AI Fire
- Posts
- 🕵️ Claude’s 151M Query Heist
In partnership with
So apparently, “stealing” an AI model may just mean asking it 151 million questions. Anthropic says Alibaba and others did exactly that. AI arms race is getting… weird?
What's on FIRE 🔥
IN PARTNERSHIP WITH AGORAPULSE
Archie is trained specifically on high-performing social content and built around your own webinars, articles, and decks.
Feed it a source, and it produces on-brand posts for LinkedIn, Instagram, and Facebook with recommended timing pulled from your own engagement data, while you stay in control of every post before it publishes.
AI INSIGHTS
Anthropic says Alibaba, Moonshot AI, and DeepSeek ran large-scale campaigns to extract Claude’s capabilities and use the outputs to improve their own models. Main findings:
Alibaba allegedly generated 151M+ suspicious Claude exchanges between May and July 2026, targeting capabilities useful for Qwen.
Moonshot reportedly routed 23M+ interactions through Claude over roughly three months.
The campaigns used thousands of accounts, proxies, aggregators, and “transfer stations” to disguise their origin.
Prompts focused on valuable skills such as coding, agentic reasoning, computer use, and tool calling.
Anthropic’s earlier investigation found 16M+ exchanges from DeepSeek, Moonshot, and MiniMax through about 24,000 fraudulent accounts.
DeepSeek allegedly used Claude to reconstruct reasoning and even grade outputs for reinforcement-learning workflows.
Distillation itself is a common AI training technique. The controversy comes from allegedly extracting another company’s outputs at industrial scale while bypassing access restrictions and terms of service.
🎁 Today's Trivia - Vote, Learn & Win!
Get a 3-month membership at AI Fire Academy (700+ AI Workflows, AI Tutorials, AI Case Studies) just by answering the poll.
What was notable about Moonshot’s alleged approach? |
PRESENTED BY ATLASSIAN
AI made PMs faster. Multiplayer mode is still broken.
A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.
Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.
And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.
AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.
AI SOURCES FROM AI FIRE
🎉 Today, let's shout out to our legends winning 3-month pass to AI Fire Academy of Aug 2026! Enjoy all your PRO club benefits, let me know if you want any guides.
elprofe****@gmail.com | leopold1w*****@gmail.com
bdow****@***erialbeverage.com | natalia****[email protected]
1. Free Guide: GPT Images 2.5 is Seriously Impressive: Every New Feature (King of AI Images?). I break down the updates one by one so you can quickly understand what changed, what’s genuinely useful, and how you can put the new features to work.
2. Copy a $1K+/Day YouTube Channel with AI from Scratch. Step-by-Step (Everything You Need). A practical guide to using AI to make money with story-driven YouTube videos, from host creation to final export. The goal is a repeatable process you can actually run yourself.
3. How to Turn Claude Into Your One-Person Marketing Team in Under 1 Hour (Part 1: Set It Up). I built a marketing system that can understand how I work and help me handle real marketing tasks. Part 1 reveals the setup that makes the whole workflow possible.
4. GPT-6 Astra or Fable 5.1? I Tested Both, If I Could Keep ONE & ONLY, I’d Pick... (No Hype). I put 2 models through 3 real build tests to answer this simple question. No marketing claims. Which model I’d choose for everyday work?
TODAY IN AI
AI HIGHLIGHTS
📊 OpenAI launched ChatGPT for Financial Services, that can plug institutional data straight into Work mode. It also supports GPT-6 Astra. Wall Street, now your turn.
🤩 Apparently everyone is living in the ’80s now thanks to ChatGPT. It's everywhere, the results are ridiculously good. Steal these prompts and make your own version.
🚨 Astra has been out for barely a week, OpenAI just temporarily paused new Pro subscriptions because Astra is hammering its infrastructure. Existing users are safe.
📱 The iPhone Duo’s foldable display is already inspiring a wave of new app concepts across our community. Devs are starting to prove there’s much more to play with.
🎧 Suno just launched v6 after rebuilding its music models with Warner, BMG, and Believe using licensed data. Fan remixes with artist payouts are also coming next.
📈 Meta’s new AI agent Muse is now the No. 2 app in the U.S. App Store with 83K+ iOS downloads. It’s not ChatGPT-level yet, but Meta clearly has people curious.
🤯 Jensen Huang says Nvidia could grow revenue another 70% next year, potentially reaching around $680B. He sounds pretty confident, and he can see the future.
💰 Big AI Investment: Google will invest $15.1B in Finland through 2028, its biggest-ever European investment. The plan covers AI data centers, energy projects, and a 22-year power deal with Fortum.
NEW EMPOWERED AI TOOLS
🌐 Gemini for Windows brings AI right to your desktop to polish drafts, brainstorm, and create images or videos with Alt + Space.
🚀 Nebius AI Builder Program gives developers $400+ in free credits, production-ready AI blueprints, and access to NVIDIA courses.
💤 Dreambeans curates personalized story collections from your connected apps and Gemini, letting you explore ideas, save favorites.
🤖 Typewise Nova builds, tests, and continuously improves AI customer support agents, handling orders, refunds, and plan changes.
AI BREAKTHROUGH
DeepSeek launched V4.1 Flash, the model has 552B total parameters, it also supports native image understanding and context windows up to 1 million tokens.
Its biggest architectural trick is the KV cache. DeepSeek says V4.1 Flash uses roughly 4× less KV cache than V4 Flash and 437× less than the original DeepSeek V1. So the “437× smaller” claim is real, but it’s specifically a comparison with DeepSeek V1, not GPT-5.6 Sol.
On DeepSeek’s own evaluations, V4.1 Flash actually beats GPT-5.6 Sol on several agentic benchmarks:
Terminal-Bench 2.1: 90.6 vs 88.8
DeepSWE v1.1: 74.2 vs 73.0
CyberGym: 88.1 vs 84.5
AutomationBench: 54.8 vs 45.8
Agent’s Last Exam: 31.8 vs 26.7
NL2Repo-Bench: 64.0 vs 56.8
DeepSeek also trained it heavily for agentic workloads using large-scale synthetic environments and reinforcement learning. Users can adjust its reasoning effort from 1 to 100, trading compute for stronger results.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply