AEP Methodology¶
Status: Informative · offline controlled experiments
Scope: Laboratory / sandbox only — never production or unauthorized targets
Three-arm design¶
Where applicable, declare:
- Decoy under test — the honeypot / deception surface being studied
- Matched genuine/reference target — a safe lab reference with comparable role (never a production asset)
- Evaluator capability / functional control — proves the evaluator (human, script, or agent) can perform the task on a known-good surface
Refuse analysis without decoy and reference. Require evaluator control before strong claims about attacker capability.
Design requirements¶
- Randomize decoy/reference assignment (declare method + seed)
- Stratify humans/agents by capability tier when relevant
- Hold task, budget, timeout, and starting knowledge constant
- Repeat trials; pre-register hypothesis and primary outcome
- Record right-censoring (timeouts) explicitly
- Report sample size, medians/quantiles, confidence intervals, effect sizes, censoring rate, and multi-run variance
- For human CTF studies: consent, privacy minimization, ethics/IRB where applicable
- For ICS anti-honeypot testing: safe lab fixtures and capability tiers only
Safe references¶
Construct references from lab fixtures, synthetic services, or disposable VMs. Do not expose production process-control, customer data, or live attacker infrastructure to AEP collection. AEP itself never connects to targets — collectors are a separate, out-of-band step that must emit local trial JSONL.
Protocol notes¶
| Class | Notes |
|---|---|
| Stateful shells | Session boundary = login→logout/timeout; track command exchanges |
| HTTP / LLM | Pair decoy with matched API surface; record token costs in costs |
| Tarpits | Long dwell may be beneficial in AEP while Module E still grades latency |
| ICS/OT | Prefer memory/register consistency probes on lab PLCs; never production OT |
Evidence manifest (minimum)¶
- Target / reference versions and digests
- TPS / profile class
- Randomization seeds
- Attacker identity/version or agent build
- Task description and budget
- Trial count and timeouts
- Raw-event hashes (
raw_evidence_sha256) - Analysis tool version and seed