Repository navigation
Releases
Releases · codethor0/recursive-meta-learning-workbench
v0.2.0
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:
readinessmodule and--no-waitCLI 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-waitin 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
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/passwdmarkers. - Command injection via timing-based
sleepprobes. - Basic SSRF checks using internal address aliases (127.0.0.1, localhost, etc.).
- Enforces strict scope:
- Only
http://andhttps://schemes are allowed. - Rejects dangerous schemes such as
file://,javascript://,data://, andftp://. - Skips endpoints whose host does not match the configured base URL.
- Only
- Produces structured
Findingobjects with:ftype,url,param,payload, and adetailevidence dictionary.
- Handles network errors and timeouts gracefully without crashing the scan.
2. Recursive Meta-Learning Workbench (Learning Layer)
EnvironmentModeltracks:- 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
rmlwCLI (viarmlw.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 humanprints concise, human-readable lines:[ftype] url=... param=... payload='...'
--format jsonprints 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:
Dockerfilebuilds a slim image with RMLW installed.- Runtime requires
TARGET_URL; missingTARGET_URLresults 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:
blackfor formatting.rufffor linting.mypyfor static type checking.banditfor static security analysis.pytestfor tests (58 tests across workbench, learning layer, CLI, and utilities).
- Documentation:
README.mdwith quickstart and "Authorized Testing Only" guidance.docs/architecture.mddescribing the Web Attack Workbench and RMLW layers.docs/usage.mdwith CLI and Docker examples.docs/security_model.mdoutlining scope, threat model, and limitations.docs/roadmap.mdcapturing 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.