What they did
David Ray Jr, a penetration tester, built Specter Defense on Kali Linux to see whether a defensive system could catch AI-style offensive activity. It combines Nmap, TShark, Python and SQLite with an AI investigator running Qwen3 14B through Ollama. Fixed rules flag suspicious traffic first; the investigator then gathers evidence with its tools and writes a structured risk assessment. A separate validation step checks the AI's conclusions against the evidence, and a policy engine decides whether to escalate. Containment needs a human to approve it. He tested it in an isolated lab network where a simulated attacker ran ARP sweeps and TCP port scans.
How it went
The rules caught three reconnaissance patterns, including high-severity ARP sweep and port-scan detections. The investigator used seven evidence tools and rated the activity high risk at 85% confidence, and its findings passed validation. The policy engine escalated to a person instead of acting alone. Once approved, filtering rules isolated the attacking machine and all six test packets were blocked; after release, all six recovery probes got through.
Worth knowing
The attack traffic was a scripted simulation, not an adaptive agent, so this shows the pieces working together rather than real-world detection rates. The author deliberately kept the model out of enforcement: it recommends, deterministic checks and a human decide. Run anything like this only on a network you own or are allowed to test.