See pricing
Rebuilding the Internet for Privacy | Barrett Lyon on DoxxNet
The a16z ShowSecurity

Rebuilding the Internet for Privacy | Barrett Lyon on DoxxNet

A veteran of DDoS defense and internet mapping explains why VPNs only fix part of the privacy problem, and how he built a peer-to-peer mesh network with 12 people, 26 sites and AI instead of sysadmins.

October 1, 2026 · 45 min listen · 10 min read · Barrett Lyon
0:00
–:––

Context

Barrett Lyon is an internet infrastructure veteran. He built early DDoS defense (Prolexic), a video-focused CDN (BigGravity) and a famous map of the whole internet. He is now building DoxxNet, a private network with built-in apps for chat, calls and file transfer. Joel de la Garza of a16z interviews him. It matters to PMs because the episode shows what a product looks like when privacy is the architecture, not a settings page, and how a tiny team can run global infrastructure with AI.

The Big Idea

Most private-messaging apps still send everything through someone else's server. DoxxNet's bet is that privacy only becomes real when you own the network, remove the middleman, and let devices talk directly.

Barrett argues that internet protocol design has stagnated, so the biggest privacy gains come from rebuilding the layer underneath the apps, not from adding features on top.

Key Insights

VPNs only solve part of it

  • What: Barrett says most VPNs are just entry and exit points. The provider leases a server from a landlord, installs software, and your traffic comes out of that node.
  • Why it matters: "No logs" claims leave a bigger question open: whose servers are those, and are they owned or leased? If they are leased, you also need to know who they are leased from and what that party does with the data.
  • His conclusion: "if it's not your equipment, you can't set policy." The next layer of policy is the carriers and interconnects beyond your own rack.

A parallel network inside the VPN

Barrett uses VPN protocols only as the door into a separate world. Inside it, DoxxNet runs its own mesh network.

  • 196 domains that do not exist on the public internet.
  • Users or agents can register domains, lease mesh or public IP addresses and set policy through an API in milliseconds.
  • Best-in-class apps (chat, voice, video, file transfer) sit on top and run across the mesh, all encrypted.
  • No phone numbers, email addresses or names are needed. You prove you are human with a small puzzle and get a token.

No server in the middle

Barrett's complaint about WhatsApp, Signal, iMessage and Telegram is that they look like you sending something to a person, but you are really sending it to a server that the other person then pulls from.

  • On DoxxNet, your phone becomes a server for your friend.
  • He says you can send a 20 gigabyte file with no file size limit.
  • Latency is point to point, with three or four layers of encryption, and he describes it as post quantum.
  • This also removes the slow upload leg to a distant central server.

Hide the cryptography from users

Joel noted that a key lesson of public key infrastructure is that if you tell people it is PKI, it will not work. DoxxNet follows that. A user just invites a friend. Behind the scenes the two devices do a cryptographic exchange, and trust between them unlocks direct calls and file transfer.

Own the whole stack

  • What: DoxxNet owns its infrastructure from software to wire, including the AI that handles routing.
  • Scale: about 26 sites around the world with redundant routing, built in roughly six months, with about 12 people involved.
  • Operations: no network administrator and no sysadmin. AI does that work.
  • Why it matters: a small team means a very small blast radius. Fewer employees have access, which cuts the insider threat Joel calls the cause of the most crippling cyber attacks.

Run the AI locally

Barrett was openly distrustful of large inference providers, so DoxxNet runs its own AI stack. Context and inference stay inside its own secure environment instead of being exported to one of the big AI companies. Equipment is watched by cameras, and nobody physically touches it.

Tracking plus cheap AI

Barrett's answer to "so what if someone knows I bought something at Target" is that it is never one purchase. Trackers across every app and site add up, get stored and get sold. Combined, they can reveal when you wake up, your income, where you shop and how often, and even big life events before your spouse knows.

Joel adds the new twist: you can now dump all of that into an LLM and ask it about a person. Barrett notes WhatsApp, Apple and others include telemetry or trackers, and that blocking them usually does not break the app.

Networks cannot do identity

Barrett describes censorship creeping into the US. Lawmakers ask the network to solve an identity problem, such as age checks, but "an IP address is not your driver's license." If age or content limits are needed, they belong in a different place in the stack.

His example of blast radius: a carrier was classified as France by a geo provider. France required blocking a hundred or so piracy sites, so the provider blocked the whole country, and the carrier got caught in that policy.

Mental Models & Frameworks

Control the layer you promise

If you promise privacy at one layer, you need to control the layers beneath it. Barrett traces the chain:

  • App: who can see the content.
  • Server: who owns the machine.
  • Hardware and colocation: who leases it to you.
  • Carriers and interconnect: who sits beyond your own cage.

Any link you do not control can set policy for you. Use this as a check when a vendor claims "no logs" or "end to end".

Aggregation beats single data points

One tracked event looks harmless. The risk comes from combining many. When reasoning about privacy features, ask what a stranger could infer from all the signals together, not whether any one signal is sensitive.

Governance as a product feature

DoxxNet is US based, with a separate Swiss company handling European operations so the two regions are separated. Barrett says corporate governance matters here, and the company is open about who owns it, who is involved and why decisions were made. He also says this helps drive community.

Decision Principles

Principle: Disclose ownership to earn trust

  • When: you sell a privacy or security product and users cannot verify your claims.
  • Why: Barrett is uncomfortable using services that do not say who owns them or why they exist beyond privacy. Being open about who is involved is part of the product.

Principle: Put each rule at its layer

  • When: a policy problem (age, region, identity) is being pushed onto the network or infrastructure.
  • Why: networks do not do identity, so rules enforced there cause collateral damage, such as a whole country being blocked.

Common Mistakes

Mistake: Trusting free or obscure VPNs

  • Why it happens: app stores list dozens of VPNs, many free.
  • Why it is harmful: Barrett and Joel say that if you are not paying, you are the product. Joel adds that in some countries intelligence apparatus and organized crime overlap, and early crypto users had accounts drained through such services.
  • Better approach: ask who owns the servers, who leases them and what the policies are.

Mistake: Trusting a niche secure app

Joel tells of organized criminals who avoided Signal, thinking it was the NSA, and used an app only a few people used. It turned out the FBI had built and marketed that app, and many of them were indicted. The lesson is that "only a few people use it" is not a security property.

Practical Application

Audit who sees your user data

  • Do: list every server, vendor and tracker that touches user messages, files or telemetry in your product.
  • Then: mark which ones you own and which are leased or third party.
  • Why it works: Barrett's rule is that you cannot set policy on equipment you do not own.

Design privacy so users skip jargon

Hide cryptography behind an action users already understand, like inviting a friend. DoxxNet's trust exchange happens in the background when you add someone.

Pilot AI-run operations

DoxxNet runs global infrastructure with no network admin or sysadmin. Pick one operations area, such as routing or provisioning, and test AI doing the routine work while humans keep policy and review.

Check classification before you launch

Barrett's team was flagged by content filters, and Joel's companion-chatbot example was labelled adult content by half the classification engines. Before launch, test your domain and product name against common filters and plan the appeal process early. DoxxNet applied through proper channels with one AI provider and got everything opened up.

Questions to Consider

  • Which of our product's promises, like "your data is private", depend on a vendor server or third party we neither own nor audit?
  • If someone combined every analytics tag in our app into one profile of a user, what life events could they infer?
  • Where are we asking one layer, such as the network, to solve a problem that belongs elsewhere, such as identity?
  • If we replaced the central server in our core feature with direct device to device connections, what would improve and what would break?

Bottom Line

Barrett Lyon argues that real privacy needs control of the whole stack and fewer intermediaries, not just a VPN on top of someone else's servers. With cheap AI turning years of tracking data into profiles, he sees rebuilding the internet's lower layers as urgent, and shows a 12 person team can start.

Concepts to Explore

Carrier-grade NAT

Barrett calls it one of the biggest stains on the internet because it destroys applications' ability to do peer-to-peer. Erasing that obstacle is a stated goal of DoxxNet's mesh design.

Mesh networking

A network where devices connect directly to each other instead of through a central hub. Barrett points to powerful Wi-Fi at the edge and older wireless packet networks as inspiration for better mesh and P2P protocols.

Case Studies Mentioned

Surviving DDoS attacks

Barrett began defending his own servers from attackers in high school, wrote down what it takes to survive an attack, and presented it at SANS. Calls followed from businesses facing extortion, such as watch sellers attacked during Christmas sales or gambling sites going dark before the Super Bowl. He built Prolexic around leasing big, expensive scale as a service, one of the first cloud services. It still exists as Akamai Prolexic.

BigGravity video CDN

Barrett built a video-focused CDN that saved customers money. His example: a progressive file download wastes bandwidth when someone just scrolls into the middle of a video.

The France blocking incident

A geo provider labelled a carrier as France, so a French order to block piracy sites led to the entire carrier being blocked. It shows the blast radius of classification errors.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
DoxxNetPrivate mesh network with built-in P2P chat, calls, video and file transferBarrett's company and the subject of the episode.
Cirrus aircraft with ballistic parachuteSmall plane with a whole-airframe parachuteBarrett's engine failure story, and why pilots hesitate to pull the chute.

People to Follow

Barrett Lyon

Founder of DoxxNet. Previously built Prolexic, BigGravity and Defense.Net, and created the Opte Project internet map that entered MoMA's permanent collection.

Joel de la Garza

a16z partner and host of the conversation, with a security background and early work on DDoS tools.

Notable Quotes

"If it's not your equipment, you can't set policy." (Barrett Lyon)

"An IP address is not your driver's license." (Barrett Lyon)

"If you're not paying for it, you are the product." (Barrett Lyon)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free