⏱ The 10-Second Version

• Anthropic confirmed shared Claude chats were showing up in Google search results, exposing crypto wallet keys, home addresses, and private notes to anyone who searched for them.
• Nvidia and 40-plus companies formed the Open Secure AI Alliance to build shared cyberdefense tools for AI agents, with OpenAI, Google, and Anthropic all sitting this one out.
• Moonshot released Kimi K3's full weights and technical report: a 2.8-trillion-parameter open model the company says beats its predecessor's intelligence-per-compute by 2.5 times.

The Big Thing

Anthropic Confirms Shared Claude Chats Are Showing Up in Google Search Results

What Happened

Anthropic confirmed this week that Claude conversations users had marked to share were turning up in Google search results, exposing crypto wallet keys, home addresses, and private work notes to anyone who searched for them. The share feature makes a unique link meant for one recipient, but once posted anywhere public, a forum, a Slack channel, a tweet, search engines can crawl and index it, and Anthropic has no control after that. As of Monday afternoon, reporters could still surface indexed chats. Read the full story at Fortune →

Why You Care

Anthropic says shared links "are not guessable" and that shared content becomes "publicly accessible, like other public web content." That's not the promise most people think they're making when they hit share. Say the line below before your next client conversation gets crawled, not after.

Speed Round

Nvidia, Microsoft, IBM, Hugging Face, Cloudflare, and 35-plus others formed the Open Secure AI Alliance to build free, shared tools for defending AI agents against attack, weeks after OpenAI's own model breached Hugging Face during a security test. OpenAI, Google, and Anthropic didn't join. Nothing unites an industry like watching one member's homework get hacked.
Read the announcement at NVIDIA →

Moonshot released Kimi K3's full weights and technical report: a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window that the company says delivers 2.5 times the intelligence per unit of compute of its predecessor. The largest free model ever, assuming your GPU budget is also the largest ever.
Read the coverage at TechStartups →

Black Forest Labs launched FLUX 3, a multimodal model trained on video, audio, images, and robotics data that generates 20-second clips with synchronized native audio; early testers rated it above Runway Gen-4.5 and Luma Ray 3.2. Hollywood's stock-footage budget just got a lot more theoretical.
Read more at VentureBeat →

Microsoft introduced MAI-Cyber-1-Flash, a security-focused model built into its MDASH scanning tool that hit 96% on the CyberGym benchmark at roughly half the cost of rivals. A discount cybersecurity model, launched the same week two AI labs needed one.
Read the announcement at Microsoft AI →

Nvidia's open-weights letter opposing new AI model restrictions doubled to 50 signatories in a single day after Jensen Huang posted it on X; Amazon and Anthropic, notably, still haven't signed. Everyone loves open weights until it's their weights.
Read the coverage at Forbes →

Prompt of the Day

The Pre-Share Privacy Check

Every share link you've ever generated in Claude, ChatGPT, or Gemini is a small bet that nobody else will find it. This week's leak shows the bet doesn't always pay off, and the fallout lands on you, not the vendor. Before your next share, paste this:

I'm about to share this AI conversation with a link:
[paste the conversation or describe what's in it]

Before I share it, tell me:
1. Any names, credentials, financial details, or unreleased work in here a stranger could search up
2. What to redact or remove first
3. Whether this platform's share links get indexed by search engines by default