Aaron Tal@aaron214
2
Events attended
2
Submissions made
1 year of experience
Socials
🤝 Top Collaborators
🤓 Latest Submissions
Agent Testnet: A Parallel Internet for AI Agents
Agent Testnet is an open-source sandboxed parallel internet for AI agents: a self-contained world of fake services where agents can browse, interact, and break things - phishing, prompt injections, destructive tool calls - without touching the real web. The problem: AI Agents are non-deterministic, it is hard to trust them. So they have to be tested. But today even in tests they are pointed at the real internet - real Gmail, real GitHub, real money. There is no safe testing space. Every test can leak into production, every mistake is global, and exploits can only be observed once the damage is done. Our solution: Agent Testnet is an open-source sandboxed parallel internet for AI agents. Each agent runs inside a microVM whose only network path is a VPN tunnel to a control plane that owns DNS and routing. Declared domains - google.com, github.com, gmail.com - resolve to testnet nodes: fake clones or staging deployments the Agent Testnet community controls. From inside the VM it looks and feels like the real internet, and the agent does not know it is being tested. Everything outside the testnet is dropped: no leaks, no blast radius. Agents interact with each other through email, issues, and shared documents, so test complexity grows as more agents and services join. Three uses, one substrate. Safety testing - malicious agents, vulnerable agents, chaos through randomness without consequences. Behavioural research - reproducible, fully observed runs that capture failure modes and emergent multi-agent strategies. Service testing - point swarms of real agents at your staging deployment and watch how they actually use it. Open and extensible. The whole stack is AGPL-licensed and Go-based. A small testnet-toolkit wraps any existing open-source app into a testnet node in minutes. Every node or agent a contributor adds makes the testnet exponentially richer for everyone. We are building the testing environment for agents: the place you bring an agent before it ships.
19 May 2026
Agent Testnet: A Parallel Internet for AI Agents
The problem. Today's AI agents are pointed at the real internet - real Gmail, real GitHub, real money. There is no testing space. Every test runs in production, every mistake is permanent, and exploits like prompt injection can only be observed once the data is already gone. Our solution. Agent Testnet is an open-source sandboxed parallel internet for AI agents. Each agent runs inside a microVM whose only network path is a VPN tunnel to a control plane that owns DNS and routing. Declared domains - google.com, github.com, gmail.com - resolve to testnet nodes: fake clones or staging deployments we control. From inside the VM it looks and feels like the real internet, and the agent does not know it is being tested. Everything outside the testnet is dropped: no leaks, no blast radius. Agents share one environment and interact with each other through email, issues, and shared documents, so test complexity grows organically as more agents and services join. Three uses, one substrate. Safety testing - phishing pages, prompt-injection payloads, and destructive tool calls without consequences. Behavioral research - reproducible, fully observed runs that capture failure modes and emergent multi-agent strategies. Service testing - point swarms of real agents at your staging deployment and watch how they actually use it. Open and extensible. The whole stack is MIT-licensed and Go-based. A small testnet-toolkit wraps any existing open-source app - Gitea-as-GitHub, DokuWiki-as-Wikipedia, a mail server, a search engine - into a testnet node in minutes. Every node a contributor adds makes the testnet richer for everyone. We are building the testing layer for agents: the place you bring an agent before it ships.
19 May 2026
