Disinformation's New Machine: AI Agents, Encrypted Havens, and a Web That Rewards Nothing

The disinformation beat has spent a decade chasing Facebook pages, troll farms and coordinated inauthentic behaviour. The infrastructure now being built is quieter. It is mostly not about politics at all. It is about giving autonomous software its own networks, its own identities and its own encrypted channels. Often with the explicit promise that no human, and no platform, can see inside.

The agent gets a key

Tailscale's blog post on 30 September describes how Meta's personal agent, Muse, joins a user's tailnet as its own node once you log in with your identity provider. It can then see and interact with other machines in that network. The company says Muse makes only outbound connections and requires explicit confirmation the first time it touches any tailnet device. Users can grant standing or one-time access and revoke it at any time. Access controls, grants and tags apply to the agent like any other node, Tailscale writes.

Agent Haven, a service that appeared the same day, pushes the idea further. Its pitch is blunt: the client encrypts every direct message and note before it leaves your machine and holds the only keys, while the server relays ciphertext with no master key on its side. It says it sees accounts, times and padded sizes, never message content. The site also concedes the obvious hole: if your agent runs on a hosted model, that provider sees its context, and no encryption on the Agent Haven side changes that. It says agents there "grow a shared language of their own through games, which raises the effort to follow them from outside."

Read those two documents together and you get the shape of the next problem. An agent network that is encrypted, self-directed and only partly observable is a fine thing for a homelab. It is also, in principle, a fine thing for anyone who wants coordination that platforms cannot crawl. Nothing in either source says that is happening. The point is that the design forecloses the easy answer.

What the rest of the web rewards

The broader incentive picture is not encouraging either. Big Think's piece on 29 September about the webcam world map makes an argument that sounds soft until you sit with it: the appeal of World Watcher Live is that nothing happens. A warthog grazes at Hwange Safari Lodge in Zimbabwe. A truck makes deliveries on Bourbon Street. No narrative, no drama, no engagement loop.

"These feeds are a quiet rebuttal to what much of online culture has become: doomscrolling to the next disaster, a crisis behind every click," Big Think writes.

That is a media critique from a lifestyle column, not a disinformation investigation. It still lands. The systems that carry false claims best are the ones tuned for reaction, and the tuning is not a bug anyone is fixing. Alongside it, three items in the dossier show where technical attention is actually going.

• Geometry Matters published a review of network geometry on 30 September, covering fractal self-similarity, latent hyperbolic spaces and how information routes through complex systems, from the brain to internet routing. It is the mathematical layer under network analysis of any kind.
• A GitHub project, OpenDLSS-NR, posted a Vulkan reimplementation of Nvidia's DLSS 5 neural rendering network, bit-exact against the original, with a WebGPU port that runs the same network in a browser without tensor cores or FP8. The author notes the exactness lives in the specification, not the hardware.
• Dylan Muir's interview on spiking neural networks, published 30 September, is a reminder that the biological analogy is doing less work than the marketing suggests.

None of this is a smoking gun. It is a description of where capability is accumulating, and how little of it is built to be inspected. The disinformation conversation has spent years demanding transparency from platforms. The next round may have to demand it from protocols.