Barrett Lyon raises $38M for doxx.net's private network for AI agents

Why it matters

doxx.net is placing a $38 million bet that AI agents will need network identities and security controls of their own. Lyon's previous infrastructure companies give him relevant experience, while the product's privacy and threat-blocking claims remain company assertions.

Barrett Lyon, the infrastructure entrepreneur behind DDoS-defense company Prolexic, announced a $38 million Series A for doxx.net on October 1st, alongside the launch of its open beta. The round was led by Andreessen Horowitz, with participation from Animo Ventures and Focal.vc. The product is Lyon's latest attempt to turn network security into infrastructure people can use directly, now with AI agents as a target user.

https://x.com/BarrettLyon/status/2105689236623986864

Lyon's record is unusually concentrated in internet infrastructure. He founded Prolexic in 2003 to protect networks from denial-of-service attacks; Akamai completed its acquisition in 2014. Lyon later founded BitGravity, XDN and Defense.net, according to a16z's investment announcement. His new pitch extends that experience from defending websites and enterprise networks to giving people and agents control over how they connect.

The doxx.net beta lets users connect devices and agents in private networks, configure tunnels and firewall rules, and communicate through direct calls, messages and file transfers. The company says its system runs on its own IP space, DNS root, certificate authority and hardware. It also says new users can start without providing an email address or phone number. Lyon's launch post invited users to try the beta and suggested they could send an agent to explain the service.

The central proposition is that an agent with access to a user's accounts and files needs network boundaries of its own. In a16z's investment post, partners Joel de la Garza and Martin Casado make the case that agents widen the amount of personal information a system can reach and the number of connections that need protection. doxx.net says agents can receive network identities and use an API, including native support for the Model Context Protocol, to help configure connections and security policies.

The phrase "without a server in the middle" needs a precise reading. doxx.net describes direct peer-to-peer calls, messages and transfers, while also operating backbone and DNS infrastructure. That distinction is relevant in a market where mesh-networking products already connect devices directly. Tailscale's documentation says its data traffic generally travels peer-to-peer; its coordination server handles keys, device discovery and network changes, and relay servers can carry traffic when a direct connection cannot be established. doxx.net's claim is therefore about its architecture and communications path, not proof that no company-operated infrastructure participates in the service.

Lyon says doxx.net has blocked more than 38 million threats since December, a company-reported figure from its closed beta. The company's homepage displays a different snapshot: more than 33 million threats blocked. Neither figure comes with a public methodology or a breakdown in the launch materials, so the total is best read as doxx.net's own telemetry claim, not an independently audited measure of protection.

The financing announcement positions the product as more than an agent-security tool. doxx.net also promotes private browsing, encrypted communication, DNS filtering and a parallel naming system. That broad scope puts it near mesh networking, VPNs and consumer privacy services, while its agent controls are the company's bid to distinguish the product as agents begin to take actions online on a user's behalf. The beta is open; the announcement did not provide pricing.

For Lyon, the $38 million round backs a founder returning to network security with a consumer-facing product and a substantially wider promise than attack mitigation: private infrastructure for people and the software agents acting for them. The test will be whether users trust a single provider to supply both the network and its security claims, and whether the agent controls prove useful outside the launch framing.