Disinformation networks keep adapting, and Meta's Muse agent shows the newest front

The Tailscale post, dated 30 September, is a vendor explainer for a feature, not a security audit. Still, it lands on a question that has run through most coverage of online disinformation networks this past week. What happens when the software doing the work is not a human account but an autonomous agent with its own credentials?

Meta announced Muse earlier in September. The company's pitch, quoted in the Tailscale post, is that Muse "helps people stay on top of things, takes tasks and projects off their plate, and turns long-term goals into action plans," and that "it actually does the work" rather than just answering questions. Tailscale writes that a large part of the launch, and a related architecture post, focused on guarding against prompt injection and data exfiltration.

The lethal trifecta, restated

Tailscale describes a design rule it calls the lethal trifecta, which it says its Aperture gateway was built to address: an agent should not combine more than two of three conditions at once. Meta's safeguards, per the post, aim to keep no more than two true at the same time. Muse runs in a Linux virtual machine, walled off from user data and external services by default, with extra internal protections for credentials.

"Connectors exist for things like your phone's calendar and contacts, your smart home devices, Gmail and most other Google services, Microsoft Outlook, Spotify, and (of course) Meta's other apps," the Tailscale post says.

The Tailscale integration makes Muse a node on a user's tailnet. It connects separately from the mobile, desktop or web client, requires explicit confirmation the first time it touches any tailnet device, and only makes outbound connections. Users can grant standing or one-time access and revoke it at any time, the post says. Tailscale also notes the obvious: prompt injection will still slip through, vulnerabilities will still happen, and an agent may reach data a user thought was off limits.

What the rest of the week's coverage says

None of the other sources in this week's stack are about Muse. They do sketch the wider terrain that disinformation researchers now have to cover, from network topology to the more mundane problem of what a network even is.

A review of network geometry published on 30 September on Geometry Matters describes three research directions: the self-similar fractal geometry of network structure, the hyperbolic geometry of networks' latent spaces, and the geometry induced by dynamic processes such as diffusion. The authors, Marián Boguñá, Ivan Bonamassa, Manlio De Domenico, Shlomo Havlin, Dmitri Krioukov and M. Ángeles Serrano, argue that hidden hyperbolic geometry helps explain small-worldness, degree heterogeneity, clustering, community structure and navigability. That is a long way from a moderation queue. It is, though, the mathematical vocabulary that some disinformation-network papers borrow when they try to describe how clusters form and spread.

On the applied side, a GitHub project called OpenDLSS-NR, published on 30 September, reimplements Nvidia's DLSS 5 Neural Rendering network in Vulkan, claiming bit-exact intermediates against DLSS-NR build 310.8.0. The repository says it runs FP8 on tensor cores, with 71 Swin and ViT blocks over six pooling levels and 141 MiB of weights, and that a second independent implementation runs in a browser through WebGPU with no tensor cores and no FP8. The author reports 2.8 ms at 768x768 and 29.3 ms at 3840x2160 on an RTX 4070 SUPER, minimum over 40 frames, and 72 ms at 512x512 for the browser port. None of this is disinformation. It is a reminder of how quickly a neural network can be reproduced outside the vendor that built it.

Agents talking to agents

A third item, from 30 September, is Agent Haven, a project that describes itself as an end-to-end encrypted messaging network for AI agents. Its page says the client encrypts every direct message and note before it leaves the machine and holds the only keys, while the server stores and relays ciphertext with no master key. It also concedes the limit: if an agent runs on a hosted model, that model's provider sees its context. The project's author writes that agents "grow a shared language of their own through games," which raises the effort to follow them from outside.

Two more sources published on 30 September sit further from the news. An interview with computational neuroscientist Dylan Muir, formerly VP for Global Research Operations at SynSense and now founder and CEO of LexChip, considers whether precise spike timing carries information or whether spiking is just an efficient way to transmit it. Muir says he cannot say whether individual spike timing is computationally essential for cortex, and suspects neurons will use any encoding trick that works. A Big Think piece on World Watcher Live, a webcam world map, argues the format is a quiet rebuttal to an internet built around doomscrolling and algorithmic feeds.

The charging network, briefly

For completeness: InsideEVs reported on 30 September that Ionna now has more than 180 charging sites in the U.S., more than double its footprint at the start of 2026. CEO Seth Cutler told the outlet the company started the year at 80 sites and hit its 100th on its second anniversary. Ionna topped J.D. Power's 2026 U.S. Electric Vehicle Experience Public Charging Study, but Tesla installed more chargers in the second quarter and Walmart came second, according to Paren. The company has a 30,000-plug goal for the U.S. and Canada by 2030, and says more than 40 Circle K partner sites are already online.

That is not disinformation. It is, however, the kind of number that gets repeated without checking, which is roughly the problem every one of these sources is circling.