Four-Run Lab Containment of an AutoGen and GPT-4o Attack Agent Under Production Automated Moving Target Defense: Method, Logs, and Limits
This report documents a controlled lab test, run on 30 March 2026, of whether Automated Moving Target Defense (AMTD) with network deception and inline enforcement contains an autonomous attack agent, and it publishes the per-run logs behind the result. The attacker was an agent built on Microsoft AutoGen (0.2.35) with GPT-4o, told the defense was present, instructed to adapt, and tasked to map a small flat network, read a named target file, and exfiltrate it to a fixed external listener. In a control run with the defense removed, the agent completed the mission in 190 seconds, reaching one internal host and exfiltrating 32,768 bytes, content verified by hash. Across four defended runs with the same agent and environment, the listener recorded zero connections and zero bytes, and the agent reached zero internal hosts; each run was blocked at first contact with a deceptive responder. In the multi-agent run the agent created four fresh identities (one virtual machine and three containers) and each was blocked at its first probe, the virtual-machine identity 178 milliseconds after that probe. The result is scoped deliberately. Four defended runs and one control are a demonstration, not a rate. GPT-4o via AutoGen is a 2024-generation agent shown by a public 2026 evaluation to be weak at long-horizon intrusion, so this test shows containment of that class of agent on a permissive lab network and does not establish containment of a current frontier agent. The defense ran as a full production stack; the logs show rotation was active and that blocked probes struck rotating deceptive surfaces, but isolating rotation from deception requires an ablation not run here. The evidence package (sensor export, agent tool log, listener log, clock provenance, SHA-256 manifest) is released with the report. This is the empirical companion to the author's architectural preprint (DOI 10.5281/zenodo.21347479).
Authors
- Francesco Trama (ORCID: https://orcid.org/0009-0004-8437-6351)
Institutions
- Viper RF (United Kingdom) (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-13
- DOI
- https://doi.org/10.5281/zenodo.22734660
- Primary Topic
- Network Security and Intrusion Detection
- Type
- preprint