Meta's Muse let strangers read your shared AI chats
Issue #063 · 4 min read
Meta's Muse let strangers read your shared AI chats
A sharing feature leaked personal and business secrets. Plus: Anthropic's own cloud partners are starting to compete with it.
The big story
Meta built a share button for its Muse AI chats. It also shared things nobody meant to share.
Researchers found that conversations shared through Muse exposed users' personal and business secrets to anyone who found the link, according to Ars Technica.
Meta has paused the sharing feature while it investigates — that's The Verge.
Think about what people actually type into a work chatbot. Draft contracts. Customer names. A pricing strategy they're not ready to announce.
A "share" button on that kind of tool isn't a convenience feature. It's a second product decision, with its own privacy review, that just happens to live one click away from the first.
If your team ships anything with a share-this-conversation option, this is the week to go check what "shared" actually means in your own product — does it include everything typed, or just the final answer, and who can see it once the link exists.
What shipped
GitHub now lets you convert every branch protection rule in a repo to rulesets in one action, instead of migrating them one at a time — that's the GitHub changelog. Small, but if you've been putting off a rulesets migration because it meant touching dozens of branches by hand, the excuse is gone.
GitHub also put a package firewall into public preview, aimed at malware that steals credentials the moment a package installs — before any of its code even runs (GitHub changelog). Worth a look if your org has ever had a CI box get popped by a bad npm install.
OpenAI is buying Orchestra Bio to fold into Rosalind, its life-sciences research model, and the related Rosalind Workbench tool that helps researchers plan and analyze experiments — details from OpenAI and more context from TechCrunch. It's a vertical bet, not a platform one — a reminder that the foundation-model companies are starting to buy their way into specific industries rather than waiting for those industries to build on top of them.
One to flag rather than celebrate: Anthropic's cloud partners — the same companies renting it compute — increasingly see it as a competitor, and it's reportedly struggling to secure enough capacity ahead of a planned November IPO, per TechCrunch. If you're betting a roadmap on a specific model vendor's uptime or pricing, this is a reason to keep a second option warm.
What I'd actually do this week
- Audit any "share" or "export" feature on an AI tool you own. Open one yourself, share it the way a user would, and read what actually went out. Muse shows how easily the answer is "more than you thought."
- If you rely on Anthropic's API for anything customer-facing, check your fallback plan. Compute fights between model makers and cloud providers aren't your problem until suddenly they are.
- If your repo migration to GitHub rulesets has been a backlog item, do it this week — the bulk conversion removes the main excuse not to.
Hit reply if you've had a scare with a "share this chat" feature — I'd like to hear it.