Skip to content

Architecture

Isaac Thor edited this page Mar 8, 2026 · 1 revision

Architecture

Overview

RMLW has two main layers: the Web Attack Workbench (static foundation) and the Recursive Meta-Learning Layer (self-improving orchestration).

Web Attack Workbench

The workbench runs vulnerability tests against HTTP endpoints.

Test Families

Test Signal Heuristic
XSS Payload reflection Payload appears in response body
SQLi Boolean logic True/false conditions produce different responses
LFI File content /etc/passwd markers in response
CMDi Timing sleep N causes measurable delay
SSRF Internal access Non-error response to internal IP probes

Data Model

  • Endpoint: (method, url, params, headers, body)
  • Parameter: (name, location, baseline_value) – query, path, header, or body
  • Finding: (ftype, url, param, payload, detail)

Parameters are extracted from the query string. For URLs without query params, a default "q" is used.

Learning Layer

EnvironmentModel

Captures target characteristics:

  • waf_detected: WAF presence from headers
  • input_filtering: Per-parameter filter behaviour
  • successful_payloads: Payloads that triggered findings
  • failed_payloads: Payloads that did not

RLPathOptimizer

  • State: Coarse encoding from EnvironmentModel
  • Actions: Test families (xss, sqli, lfi, cmdi, ssrf)
  • Algorithm: Tabular Q-learning

EvolutionaryPayloadGenerator

  • Population: Payload strings
  • Fitness: Based on findings
  • Operators: Mutation (encoding, case, quotes), crossover

PersistentArchive

  • Stores successful payloads with context
  • Supports recent(n) and diversity retrieval
  • Feeds recombination and novelty search

MetaController

  • UCB Bandit: Chooses which mechanism to apply
  • Strategies: RL-driven, evolution, archive, baseline
  • Reward: New findings, novelty

Runtime Flow

  1. MetaController selects mechanism (UCB)
  2. Mechanism influences Workbench (payloads, test order)
  3. Workbench runs tests against target
  4. Findings update EnvironmentModel and Archive
  5. RL/Evolution update based on reward
  6. Repeat

URL Utilities

Target URLs are validated and scoped:

  • validate_target_url: Rejects disallowed schemes (file, javascript, data, etc.)
  • normalize_url: Structural normalization
  • url_in_scope: Ensures endpoints match the configured host

All scans are restricted to the host of the configured base URL.

Clone this wiki locally