Skip to content

Releases

v0.2.0

Choose a tag to compare

@codethor0 codethor0 released this 08 Mar 22:09
85917f7

v0.2.0

Learning lab harness, Cursor master prompts, docs polish, and readiness check.

Highlights

  • DVWA lab: scripts/run-dvwa-lab.sh (build, bring up DVWA, baseline + learn run, tear down); scripts/pre-push-check.sh; getting started for contributors in CONTRIBUTING.md
  • Readiness: readiness module and --no-wait CLI option; CLI waits for target HTTP 200 by default
  • Learning docs: learning-lab-dvwa.md, learning-evidence.md, learning-proof-run.md (proof run recipe and log-line guide)
  • Cursor prompts: Master prompt from-scratch and continue; daily-driver prompt; learning-lab instructions (all in docs/)
  • Code: MetaController reward docstring and reward-is-finding-count test; CLI human-format exact-pattern test; usage.md --no-wait in options
  • Master prompt: Gaps-not-rebuild guardrail, quality gate block, redirect-out-of-scope behavior, align-existing-files in PHASE 1

Quality

black, ruff, mypy, bandit, pytest (65 tests) all pass.

v0.1.0 – Initial Proof-of-Concept Release

Choose a tag to compare

@codethor0 codethor0 released this 08 Mar 00:11
62b53f3

RMLW v0.1.0 – Initial Proof-of-Concept Release

This is the initial public proof-of-concept release of the Recursive Meta-Learning Workbench (RMLW): a transparent, self-improving framework for web application security testing in authorized lab environments.

Highlights

1. Web Attack Workbench (Static Layer)

  • Models a target as a set of HTTP endpoints with parameters.
  • Implements targeted tests for core vulnerability classes:
    • Cross-Site Scripting (XSS) via reflection checks.
    • Boolean-based SQL injection.
    • Local file inclusion / directory traversal using /etc/passwd markers.
    • Command injection via timing-based sleep probes.
    • Basic SSRF checks using internal address aliases (127.0.0.1, localhost, etc.).
  • Enforces strict scope:
    • Only http:// and https:// schemes are allowed.
    • Rejects dangerous schemes such as file://, javascript://, data://, and ftp://.
    • Skips endpoints whose host does not match the configured base URL.
  • Produces structured Finding objects with:
    • ftype, url, param, payload, and a detail evidence dictionary.
  • Handles network errors and timeouts gracefully without crashing the scan.

2. Recursive Meta-Learning Workbench (Learning Layer)

  • EnvironmentModel tracks:
    • WAF hints from response headers.
    • Simple input filtering behavior (e.g., blocking, sanitization).
    • Successful and failed payloads per vulnerability type.
  • RLPathOptimizer:
    • Tabular Q-learning for choosing which test families to prioritize.
    • Epsilon-greedy policy with unit tests covering selection and updates.
  • EvolutionaryPayloadGenerator:
    • Maintains a population of payloads with fitness scores.
    • Supports mutation and crossover operators.
    • Accepts an optional RNG for deterministic testing.
  • PersistentArchive:
    • Stores successful payloads plus context and timestamps.
    • Exposes recent() for retrieving recent successes.
  • MetaController:
    • Implements a simple multi-armed bandit (UCB) over strategies (baseline, RL-focused, evolution-focused, archive reuse).
    • Orchestrates calls into the workbench and learning components.
    • Logs strategy decisions, rewards, and basic metrics.

3. CLI and Docker

  • rmlw CLI (via rmlw.cli.main:main):
    • rmlw scan --target URL --mode baseline|learn --iterations N --format human|json --output FILE
    • Validates and normalizes the target URL before scanning.
    • --format human prints concise, human-readable lines:
      • [ftype] url=... param=... payload='...'
    • --format json prints a JSON array of findings to stdout and, optionally, to --output.
    • Exits with a clear error and non-zero status for invalid URLs or fatal issues.
  • Docker:
    • Dockerfile builds a slim image with RMLW installed.
    • Runtime requires TARGET_URL; missing TARGET_URL results in a clear usage message and exit.
    • Supports running baseline scans or short learning runs in containers.
    • Documentation includes an example workflow with DVWA or similar lab targets.

4. Security, Quality, and Tooling

  • Strict scope enforcement and URL validation utilities (url_utils.py).
  • No use of eval, exec, or unsafe deserialization.
  • No subprocess-based target interaction; all probing is via HTTP requests with explicit timeouts.
  • No emojis and no prompt artifacts committed to the repository.
  • Tooling and CI:
    • black for formatting.
    • ruff for linting.
    • mypy for static type checking.
    • bandit for static security analysis.
    • pytest for tests (58 tests across workbench, learning layer, CLI, and utilities).
  • Documentation:
    • README.md with quickstart and "Authorized Testing Only" guidance.
    • docs/architecture.md describing the Web Attack Workbench and RMLW layers.
    • docs/usage.md with CLI and Docker examples.
    • docs/security_model.md outlining scope, threat model, and limitations.
    • docs/roadmap.md capturing future directions (richer discovery, more vuln classes, LLM/Fisher-Rao extensions, and CI/CD integration).

Intended Use

RMLW v0.1.0 is a research and training tool. It is intended only for use against systems and applications where you have explicit, written permission to test. It is not a point-and-shoot scanner and is not designed for unauthorized hacking or internet-wide scanning.