// research log

Agentic AI for autonomous VAPT in operational technology.

My doctorate at Keele University asks whether autonomous agents — not scripted scanners, not LLM chat wrappers, but goal-directed systems that observe, reason, decide, and act — can assist with vulnerability assessment in environments where downtime kills people.

// pathfinder ai

Pathfinder AI

An agentic vulnerability assessment framework for operational technology environments. Built around the PEER lifecycle (Plan → Enumerate → Exploit → Report) with an inner ORDA control loop, governance-gated tool calls, and hash-chained telemetry. Evaluated against a SWaT-inspired multi-PLC Modbus testbed.

Status
Phases 0–3 complete · end-to-end verified · 142 tests passing
Stack
Python 3.11+ · Ollama (DeepSeek R1 14B) · MCP · Jinja2 · async httpx
Testbed
SWaT-inspired multi-PLC Modbus environment
Scope
Two zones · three vulnerabilities · five benchmark runs
Repository
GitHub Repository →
In depth
Full project page →

OT environments don't tolerate the assumptions IT pentesting tools were built on. You can't just nmap a PLC and walk away — the device may behave correctly, the network may not survive. Autonomous tooling has to incorporate caution as a first-class concern, not a feature flag.

Pathfinder AI is deliberately scoped down. Single agent, not a multi-agent orchestration. Local models, not API-tethered. A defensible minimum experiment, with a pre-written viva argument for why "small and demonstrated" beats "ambitious and unfinished".