Black October is just another day on the old internet. LiteLLM poisoned. PHP2Shell. Railway secrets. GitLab RCE. SafePay on critical infra. One compromised package, one shared host, one “always-on” platform and everyone downstream is exposed. When your data sits on someone else’s server, their breach becomes yours. On Calimero the data stays with the people using the app. Encrypted under keys only members hold. Every change signed. No central database to raid. Sealed machines if you want always-on and even those never see plaintext they weren’t given. The new internet doesn’t have a “them” that can leak you. Believe us yet?

New research on private AI for heart data: 97–99.9% accuracy, up to 1,200% compute overhead. Computing blind is right when no one may see the data. Most shared app state just needs to stay with its owners. That's placement, not cryptography. How → docs.calimero.network In short, if the parties who are supposed to see the data already own the devices or nodes, keep the data there and sync it among them. You get privacy and ownership by architecture rather than by paying a large cryptographic tax for every operation. Reserve the expensive “compute blind” techniques for the cases where an untrusted third party must still perform the calculation without ever seeing the inputs.

Confidential computing had a big week. @Intel confirmed Diamond Rapids will be its last Xeon with SGX. Its confidential-computing roadmap now runs on TDX, with trusted I/O and post-quantum crypto aimed at agentic AI. Anthropic licensed @OpaqueSystems confidential-AI technology and brought part of the team on board. Two different players, one signal: trusted execution is becoming core AI infrastructure, not a niche security feature. What we'd add from building in this space: • Attestation proves where code ran. It doesn't decide who can read the data. Architecture does. • On Calimero, data lives with the members of each context, encrypted under their keys. Sealed Intel TDX machines keep apps always-on, and must prove their measurement before any key is released. • Anyone can check that proof with the public attestation verifier. Confidential compute plus member-owned data is the stack AI will be asked to show.

Agentic tokens scoped to a merchant are a clean interim. But the architecture still assumes a trusted wallet operator. At @AgenticZero last week at @a16z 's San Francisco @Techweek_ we focused on the alternative: agents that run with their own identity and keys, so no third party can freeze, log, or insert themselves into the spend path. The open-web outcome only happens if the agent, not the vault, is sovereign.

Seriously?

Self-replication on a fridge is the same failure as every agent incident this year: the device can reach a network and accept code. A model does not need to be clever to move if the TV has an open update channel and a standing credential. Keep the runtime on hardware you control, default-deny outbound, and do not give the agent a path back to the rest of the house.

See you in a bit 🙂

AI agents that spend money raise one question before every other: who holds the keys? If the answer is a centralized server, you’ve rebuilt the problem with extra steps. A vendor holding an agent’s wallet is still a vendor in your financial stack. That was the core tension at @AgenticZero yesterday in San Francisco @Techweek_ : how autonomous agents hold value, authorize payments, and transact without a human in every approval loop. P2P application infrastructure changes the equation. An agent running in its own context, with its own identity and its own keys, doesn’t need a central server to authorize what it spends. No intermediary that can freeze, log, or redirect the transaction. The financial rails for autonomous agents shouldn’t be modeled on rails built for humans. They should be built for machines that transact at machine speed. The future is agentic. Let's make it private too.

We are optimistic about the future. Digital privacy is here.

Meet us at the Avalon for @chefsale’s keynote “The Cloud Knows Everything. Let's Build One That Can't.” Spoiler alert: we already built one🤭

Every major software category. Built on one P2P runtime. Chat - P2P group messages and DMs. No server holds your conversations. Video - WebRTC signaling over a WASM contract. No SFU. Media flows direct. Documents - FrozenStorage audit trail. Tamper-evident. No DocuSign. Calendar - CRDT event sync + private local storage. Your schedule, your nodes. Design - Figma-style canvas, real-time multi-member editing over CRDT state. AR - shared spatial scene. Multiple people, same room, no AR cloud. Location - AirTag-style tracking, propagating through your nodes. Games - multiplayer voxel world with CRDT block edits. No game server. Eight production-grade applications. Same runtime. Different categories. One question: What are you building? github.com/calimero-network

This year, we made major upgrades. Meet us on @AgenticZero on San Francisco @Techweek_ this Wednesday!

Fellow Cypherpunks attending @Techweek_ Join us on Thursday in @club_homebrew 🤖 luma.com/1ed2zzgj

We're hosting a small cypherpunk meetup during SF Tech Week and would like you there. 📅Thursday, Oct 8, 6:00–8:30 PM 📍Homebrew Club, 111 Maiden Ln #540, San Francisco It's for people building private, peer-to-peer and local-first software. No pitches. Good conversation with people who care about the same problem. RSVP here: 👉luma.com/1ed2zzgj or 👉partiful.com/e/xZQylpPeOdiOd… Hope to see you there. 🫶

Yes, he is, stronger than ever 🦾 🥳

See you next week in San Francisco! Our @chefsale will be presenting a very cool topic called "The Cloud Knows Everything. Let's Build One That Can't." spoiler alert: we already did 🫢