Info-stealers are hunting agent folders now — tokens, MCP configs, chat history.
If your agent can read the key, so can the malware.
Give agents access. Never the keys.
@zynap_ai Holds credentials + acts on systems you answer for = the whole problem. Scope what it can do, and make sure you can yank that access without rotating every provider.
@ranjankumar Kill switch that actually stops the next tool call > hoping the agent notices it’s looping. If you can’t revoke access in one click, you don’t have a kill switch — you have a blog post.
@KOTOBUKIblogs The worst part is the UI saying “connected” while you’re on a dead token. Three secrets in play and one green check is how people lose a morning.
@AIgossipTalks This. Intelligence is cheap now. Permissioning is the product. If the agent can reach MCP + terminal + DB with one fat key, you don’t have an agent — you have a blast radius.
Built Agent Master Key for exactly this: local vault on Mac, one revocable scoped key per agent, approvals + kill switch. Free public preview (Apple Silicon).
https://t.co/yeWcgaQA8Y
Your agent doesn’t need your OpenAI key, your GitHub PAT, and your Stripe secret living in the same config like roommates who share a toothbrush.
Env vars in mcp.json are table stakes. They’re not a kill switch.
One scoped key per agent. Revoke that agent — or pause every agent — without rotating half your provider dashboard.
If your “security model” is hope + .gitignore, you’re one curious transcript away from a bad week.
Curious how other Mac agent users are handling this.
@scottshapiro “Wired to my accounts and API keys” is powerful and terrifying in the same sentence.
Recipe is fine. Custody of those keys is the part most recipes skip.
@noir_nero2025 Accidentally pasting a key into the wrong repo is a rite of passage nobody asked for.
The fix isn’t “be more careful.” It’s architecture that assumes you will mess up.