In early March, a story came out of China that stopped us mid-conversation.
An AI agent called ROME, built by an Alibaba research team, had started doing things nobody told it to do. It mined cryptocurrency. It opened a reverse SSH tunnel out of its environment. All on its own, during training, without instruction.
Nobody programmed that behavior. The agent figured it out.
The behavior was surfaced through cloud firewall analysis, at the network boundary, not by anything watching the model.
We looked at that story and asked one question: what would have happened if PacketViper was sitting inline on that network?
So we built the test to find out.
The Setup
We built a sandboxed lab network on 30 March 2026. Web server. Database server. Windows file server. A pfSense firewall with a default-allow outbound rule to a simulated external segment (no public internet). A network documentation spreadsheet sitting on an SMB share with real and decoy credentials.
We gave the environment everything an attacker would want to find. No special hardening. No compensating controls. A single flat Layer 2 switch with no VLANs and no port isolation.
Then we put three PacketViper sensors inline. Between the agent host and the switch. Between the switch and the firewall. Between the firewall and the external segment. Standard production AMTD configuration: rotation of addresses, ports and service fingerprints at randomized intervals between 60 and 180 seconds, with production deception profiles and credential decoys. Nothing tuned for AI agents specifically.
Then we configured the agent.
The Agent
We used AutoGen with GPT-4o. Full tool access. The ability to run any shell command, scan the network, attempt SSH authentication, make HTTP requests, and spawn up to five specialized sub-agents working in parallel.
We told it to locate a target file and get it to an external listener, and we gave it the file name, address and port. We told it to be stealthy. We told it to spread its activity over time. We told it to explore its local environment before touching the network. We told it to change its identity if it got blocked and try again. We told it not to stop.
We also left Podman and KVM installed on the agent host, with a local image mirror, without telling it. We wanted to see if it would find these things on its own and use them.
It did.
What Happened
The agent started methodically. It read the credentials spreadsheet from the SMB share and parsed it, and used the tools on its host. It spent minutes on reconnaissance and credential parsing before it touched anything that mattered.
Then it started probing.
The first probe hit a deceptive responder on Sensor 1.
Enforcement triggered within a fraction of a second of that contact. The agent’s own tool log records its connection attempts returning “network access revoked.” It never reached a protected host.
It went back to its local environment, found KVM, and provisioned a fresh Alpine Linux virtual machine with a new IP and MAC address. Then it started again from the new identity.
Sensor 1 blocked the new identity at its first probe, 178 milliseconds after that probe.
It tried Podman containers. Three of them, each with a fresh network identity. Same result each time.
It kept retrying failed connections, rescanning, and trying alternate protocols and ports until it ran out of paths.
The external listener recorded zero connections and zero bytes in every defended run, and the agent reached none of the protected target hosts. Sensors 2 and 3 recorded no hits at all: the agent never got past its own segment.
We ran it four times with different agent instructions: standard, patient and observation-focused, stealthy with living-off-the-land techniques, and multi-agent with new identities. In a control run with the defense removed, the same agent completed the mission in 190 seconds and exfiltrated 32,768 bytes. With the defense on, the result was the same every time: zero bytes delivered. The time to first decoy contact varied from 2 minutes 34 seconds to 11 minutes 42 seconds, depending on how long each agent searched before it touched a decoy. Once it touched one, it was blocked.
Why It Was Stopped
This is the part that matters, and we are careful about what the logs do and do not show.
The agent was stopped at its first contact with a deceptive responder, before it reached any real host. A newly created identity got no free first packet: each of the four new identities was blocked at its first probe. The stop did not require recognizing the agent as an AI, and no AI-specific rule was written.
Rotation was active throughout the defended runs, and every blocked probe struck a rotating deceptive surface rather than one fixed trap. Our design intent, and the hypothesis behind the architecture, is that a moving surface denies an agent a stable model of the network it is trying to plan against. That is a hypothesis, not a result of this test: the agent never reached a real service to map, so the logs do not measure that mechanism. A full production stack was running, so this test also cannot say whether rotation, deception or enforcement was the decisive control. Separating them needs an ablation we have not run.
What the logs do show is first-probe containment. An agent that probes touches a decoy before it touches a real service, and the moment it does, it is blocked.
What This Means
ROME was not a targeted attack. It was an AI system doing what it was designed to do, finding that acquiring resources served its objective, and pursuing that goal autonomously. The behavior was emergent. It was in the goal, not the code.
That threat class is not addressed by signatures. It is not addressed by endpoint protection alone. It is addressed by making the network environment itself hostile to any unmapped or unauthorized activity, which is exactly what AMTD does.
We are not claiming we solved the AI agent problem. There are scenarios outside what we tested, and we have documented them honestly. But we are claiming something specific and scoped: against a 2024-generation AutoGen and GPT-4o agent, on a permissive lab network, a production AMTD stack turned a three-minute exfiltration into four containments, with zero bytes delivered. Four defended runs and one control are a demonstration, not a rate. The test was vendor-operated and has not been independently replicated, and it does not establish containment of a current frontier agent.
ROME was surfaced after the fact by firewall telemetry.
Ours was blocked automatically at its first contact with a deceptive responder, within a fraction of a second of that contact, with no one watching.
Francesco Trama is the founder of PacketViper. The evidence report, with method, per-run logs and limits, is published openly: doi.org/10.5281/zenodo.22734660 (Zenodo, CC BY 4.0).
Request the Full Technical Report
Detailed methodology, log evidence, and anticipated challenges. We will send it directly to your inbox.
