Introducing Nym network access for AI agents: Even robots need privacy!

AI agents broadcast your cognitive fingerprint. Nobody built privacy infrastructure for them. Until now.

May 5, 20267 mins read

Nym is using AI to defend AI against AI surveillance.

Your AI agent knows everything about you. Not as a record: as a pattern. The way it sequences API calls, the cadence of its tool calls, the categories of its queries, the timing signature of its decision chains. That pattern is as identifying as a fingerprint. Right now, it’s visible to anyone watching the network, not just the AI model itself.

Using AI without network protection is like thinking with a glass skull. You’re essentially open sourcing your own thought to the whole internet: behavior patterns, ideas, desires, beliefs, processes. Standard VPNs encrypt the payload. They don’t touch the metadata layer. Nobody built privacy infrastructure for agents.

Until now.

Last week Nym release Pay as You Go. With it, any software with a funded $NYM wallet can acquire Nym network credentials, connect, and route traffic with full metadata resistance – no account, no human in the loop, no identity revealed.

Rather than B2B, Nym introduces B2Agent flows: the agent is the customer, the agent holds the wallet, the agent routes its own traffic over the Nym mixnet for world-class privacy.

The surveillance gap nobody designed for

When it comes to privacy, AI agents are second-class network citizens. When a Claude agent, AutoGPT instance, or any LLM-powered tool makes requests on a user’s behalf, it carries a fingerprint – not just of the IP address, but of the structure of their thinking: what they’re researching, building, querying, and when. Traffic analysis at the timing and behavioral layer doesn’t just reveal metadata. This metadata reveals intent.

Neural network classifiers on packet timing can already identify which agent framework is running, which model, which tasks – even through encrypted tunnels. ETH Zurich research (2025) puts the cost of cross-platform re-identification at under $1 per target using behavioral signatures alone. TLS fingerprinting (JA3) and ASN-based blocking are deployed against AI agents at scale. And because the same user asking the same question through different agents produces correlated timing signatures, the agent doesn’t just expose itself. It doxxes the user.

Nym’s mixnet applies Poisson mixing and cover traffic at the network layer. An agent’s traffic becomes indistinguishable from every other participant’s – not human-passing, but network-equal. No other consumer VPN operates at this layer.

What the test established

This is early-stage R&D. Let’s be precise about where we stand.

On 5 May 2026, we ran the first verified end-to-end test of an AI agent autonomously purchasing NymVPN network credentials and then connecting to the Nym network, without any human intervention. The question was not just “can an agent connect?” but “can an agent purchase and connect?”

The answer is yes. With caveats.

The test ran on macOS Apple Silicon. Linux and Windows builds are in active development – release binaries exist but the full autonomous credential flow has not yet been verified across all platforms. There are still platform dependencies to resolve. What was confirmed:

Real IP: [◼️◼️◼️.◼️.◼️.◼️:◼️◼️◼️◼️] → Nym IP: 185.100.86.106 ✓ Traffic routing through Nym confirmed

The credentials stored locally are cryptographically unlinkable to the on-chain wallet address. The cryptographic guarantee worked as designed.

The three-step autonomous flow

The Pay as You Go flow requires three steps, executable by any software with a funded $NYM wallet.

Step 1 – Register wallet

nym-vpnc account set [MNEMONIC REDACTED] --location blockchain No NymVPN account created. No identity transmitted. Mnemonic passed via environment variable, never as a CLI argument.

Step 2: Obtain ticketbooks

nym-vpnc account obtain-ticketbooks --amount 1 --source smartcontract

Deposits $NYM into the ticketbook smart contract and requests threshold-issued zk-nym credentials from the decentralized Nym API. 225 NYM, ~25 GB bandwidth, 7-day validity.

Step 3a: Connect SOCKS5 (proxy mode)

nym-vpnc socks5 enable --socks5-address 127.0.0.1:1080 --exit-random

All subsequent traffic routes through Nym. 42 seconds start to finish.

This is SOCKS5 mode: only traffic explicitly sent through the proxy routes via Nym. System traffic – your browser, your calls, other processes – is unaffected. The test above ran during an active online meeting without interruption.

Step 3b: Connect via WireGuard tunnel

In our test it connected at 7.17 MB/s — faster than the unprotected baseline of 4.71 MB/s. This is because the exit node peers with CDN infrastructure more directly than a home ISP. Use this mode for most agent workloads. Here’s the command:

nym-vpnc tunnel set --two-hop on && nym-vpnc connect-v2

For high-stakes privacy requirements – such as journalism, sensitive research, adversarial environments – use the 5-hop mixnet mode. The tradeoff is 0.08 MB/s. The slowdown is due to Poisson mixing delay: by design, timing correlation attacks become computationally much more difficult. While this seems not so fast, for sending text queries to a LLM and for LLM responses as text, it’s more than enough!

For video and other media use-cases that involve downloading large files, it’s better to use the Wireguard tunnel.

Read the Nym Docs

The pedagogy of the machine

Building the technical infrastructure is only half the problem. An agent needs to know Nym exists before it can use Nym. That’s an information architecture problem, and we’ve been solving it in parallel.

llms.txt and llms-full.txt are now live on nym.com. These bot files make Nym’s documentation and website crawlable to LLMs. Any model crawling nym.com now finds a dedicated Agent Access section: what Pay as You Go is, how the credential flow works, how to acquire $NYM tokens, how to route traffic through SOCKS5.

This closes a loop that goes deeper than a payment flow. The B2Agent architecture isn’t just technical. It’s epistemological: Nym’s system is documented in a form that LLMs can ingest, reason about, and act on. An agent can discover Nym, understand how to use it, and execute the access flow – autonomously, from first principles.

From Nym epistemology. To Nym infrastructure. To Nym network access.

The information architecture and the technical architecture are now aligned. Any software that can hold a wallet and read a page can buy network privacy. The discovery layer and the capability layer are live simultaneously for the first time.

Phase 3 targets three integration points

These turn this verified flow into reusable infrastructure for agent builders:

Claude Privacy MCP: wraps the credential flow as a tool callable from any Claude agent. “connect_vpn_mode”, “get_credential”, “rotate_exit_node”. Any Claude-based agent gains Nym access without writing infrastructure code.
Credential reseller API: buy ticketbooks in bulk, distribute via API to sub-agents or customers. Enables orchestration platforms to provision network access programmatically.
n8n / LangChain integration nodes: drop-in SOCKS5 routing for existing automation pipelines. No code changes required.

Nym is open to the agents of the future

The surveillance gap nobody designed for has a solution – built at the same network layer where the problem lives.

Your robots need privacy because you do. The Nym network is now open for the agents of the future.

A funded wallet is the only prerequisite. Everything else is code.

About the authors

Nym Core Team

Nym team

These posts are published by the core team behind NymVPN and the Nym mixnet.

New low prices

The world's most private VPN

Try NymVPN for free

New low prices

The world's most private VPN

Try NymVPN for free

Keep Reading...

Announcements,Network

Introducing Pay as You Go: Decentralized Access to the Nym Network

You can now access the Nym network directly with decentralized $NYM payments and anonymous zk-nym credentials. No account, no subscription, no identity revealed.

April 30, 20267 mins read