🥊 Anthropic accuses Chinese Labs of AI Fraud

Anthropic says three Chinese labs used thousands of fake accounts to secretly tap into Claude, racking up millions of prompt exchanges in what it calls a massive distillation attack. The fallout could reshape how AI companies protect their models. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Anthropic accuses Chinese labs of model distillation fraud.

Anthropic claims three Chinese AI labs conducted large-scale distillation attacks against Claude, using 24,000 fraudulent accounts to generate over 16 million prompt exchanges. Distillation is a standard method for compressing models, yet Anthropic argues competitors used it to replicate capabilities without building them from scratch. OpenAI previously raised similar accusations against DeepSeek. The debate exposes tension between open web training practices and proprietary model protection.

If Anthropic is right, “model moats” just got a lot thinner: you can pour billions into training, then a competitor tries to copy the behavior by hammering your API with scripted prompts at scale. That pushes every lab toward stricter access controls fast, think tougher account verification, sharper rate limits, more aggressive bot and behavior fingerprinting, which can add friction for normal developers too. For teams building products on top of Claude or similar models, this is a heads-up to treat LLM access like any other critical dependency: expect policy shifts with little notice, and have a fallback plan if a provider clamps down after an abuse wave.

Evolving AI: Your Guide to Building Secure AI Applications.

As AI adoption accelerates, new attack surfaces are emerging across infrastructure, supply chains, and model interfaces. Datadog’s AI Security Best Practices Guide breaks down how to secure:

This guide provides actionable strategies to help teams strengthen AI security without slowing innovation.

Free Download →

Evolving AI: Google expands Opal with autonomous app-building agents.

Google has added an autonomous agent to Opal, its vibe-coding app that lets users build mini web apps without writing code. Powered by Gemini 3 Flash, the agent can interpret text prompts, create step-by-step plans, and choose the right tools to complete tasks. It can connect to services like Google Sheets to store information across sessions, such as maintaining a shopping list. The agent is interactive, asking users for missing details and guiding next steps. Since launching in the U.S. in July 2025, Opal has expanded to more than a dozen countries and is now integrated into the Gemini web app through a visual editor.

When Opal can turn a plain-text prompt into a mini app that picks its own tools and keeps state in something familiar like Google Sheets, the day-to-day payoff is fast: fewer copy-paste routines, fewer “where did I save that list” moments, and quicker one-person automations for things like shopping trackers, content checklists, lead triage, and internal request forms. It’s a clear push to make agent-style workflows feel normal inside the apps people already open, which raises the bar for teams building in tools like Replit and Lovable.

Evolving AI: Enterprises gain private plugin marketplaces and cross-app AI workflows.

Enterprises can now create internal plugin marketplaces that centralize how teams access and manage AI workflows. Anthropic’s new Customize hub unifies plugins, skills, and connectors, with guided setup and stronger admin controls, including private repositories and per-user provisioning. Structured slash commands simplify execution for employees. New connectors link Claude with tools like Google Workspace, Docusign, and major financial platforms. Early cross-app support lets Claude move from Excel analysis to PowerPoint presentations in a single flow.

Private plugin marketplaces mean Claude can be rolled out like normal enterprise software: IT picks approved tools, teams grab what they need, and access stays controlled through connectors to systems like Google Workspace or DocuSign. The big trend is “AI agents inside your stack,” not another chat tab, and the new Customize hub plus marketplace files make that rollout repeatable across departments. OpenTelemetry tracking and structured slash-command forms make it easier to see spend, audit tool calls, and run workflows without prompt gymnastics. Pair that with Claude moving work from Excel to PowerPoint in one run, and finance, ops, HR teams get fewer copy-paste handoffs and faster turnaround on real deliverables.

AI gets smarter when your input is complete. Wispr Flow helps you think out loud and capture full context by voice, then turns that speech into a clean, structured prompt you can paste into ChatGPT, Claude, or any assistant. No more chopping up thoughts into typed paragraphs. Preserve constraints, examples, edge cases, and tone by speaking them once. The result is faster iteration, more precise outputs, and less time re-prompting. Try Wispr Flow for AI or see a 30-second demo.

Start flowing free

QUICK HITS

🎮 New Microsoft gaming chief has “no tolerance for bad AI”.

💰 Amazon to spend $12 billion in Louisiana on AI data centers.

👏 Council on AI ethics formed to balance innovation with human dignity.

🌊 Google claims it's building data centers that barely use any water.

 *partner link