---
title: Rhodawk AI DevSecOps Engine
emoji: ๐ฆ
colorFrom: indigo
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
---
[](https://python.org)
[](https://gradio.app)
[](https://openrouter.ai)
[](https://huggingface.co/spaces/Architect8999/rhodawk-ai-devops-engine)
[](https://docker.com)
[](.)
[](.)
[](.)
[](.)
[](.)
> **"The next generation of security tooling does not find known CVEs.**
> **It finds the assumptions that developers got wrong โ before attackers do."**
## ๐ Mythos-Level Upgrade
A complete blueprint for elevating Rhodawk to **Claude Mythos-class
autonomous vulnerability research** lives under [`mythos/`](mythos/) โ see
[`mythos/MYTHOS_PLAN.md`](mythos/MYTHOS_PLAN.md) for the full living plan
(multi-agent framework, probabilistic reasoning, advanced static / dynamic /
exploit tooling, RL self-improvement, new MCP servers, FastAPI
productization). Enable with `RHODAWK_MYTHOS=1` or hit the new productization
API at `POST /v1/analyze_target` (run `uvicorn mythos.api.fastapi_server:app`).
| Layer | Module |
|---|---|
| Multi-agent (Planner / Explorer / Executor) | `mythos/agents/` |
| Probabilistic hypothesis engine + attack graphs | `mythos/reasoning/` |
| Static (Tree-sitter, Joern, CodeQL, Semgrep) | `mythos/static/` |
| Dynamic (AFL++, KLEE, QEMU, Frida, GDB) | `mythos/dynamic/` |
| Exploit (Pwntools, ROPGadget, heap, privesc) | `mythos/exploit/` |
| Self-improvement (RL, MLflow, LoRA, curriculum, episodic) | `mythos/learning/` |
| New MCP servers (5ร) | `mythos/mcp/` (registered in `mcp_config.json`) |
| Productization API | `mythos/api/` |
---
## What Rhodawk Actually Is
Rhodawk is a **fully autonomous code repair and vulnerability research system**. Point it at any GitHub repository. It clones the code, runs the tests, generates fixes using state-of-the-art LLMs, passes every fix through a 7-layer security pipeline, and opens a verified pull request โ with no human involvement in the loop unless you require it.
When tests are already passing, it switches into attack mode: it autonomously generates property-based fuzz tests, discovers invariant violations, and hands the crash payloads back to itself for patching. It is a self-healing system that simultaneously acts as its own red team.
## The Full Autonomous Loop
```mermaid
flowchart TD
A([๐ฏ Target Repository]) --> B[Clone & Fingerprint\nLanguage Detection]
B --> C{Tests Passing?}
C -->|FAILING| D[๐ง Retrieve Similar Fixes\nVector Memory โ CodeBERT / MiniLM]
D --> E[โก Dispatch Aider via MCP\nLLM Fix Generation]
E --> F[โ
Re-run Test Suite\nVerification Loop]
F --> G{Fixed?}
G -->|NO โ retry| E
G -->|YES| H
C -->|PASSING| RT[๐ด Red Team CEGIS Engine\nAutonomous Attack Mode]
RT --> RTA[AST Complexity Scoring\nAttack Surface Ranking]
RTA --> RTB[Red Team LLM\nHypothesis PBT Synthesis]
RTB --> RTC[Deterministic Fuzz Loop\nCounter-Example Extraction]
RTC --> RTD[Crash Payload Found]
RTD --> E
H[๐ฌ SAST Gate\nBandit + 16-Pattern Secret Scanner]
H --> I[๐ Supply Chain Gate\npip-audit + Typosquatting Detection]
I --> J[โ๏ธ Z3 Formal Verification\nInteger Overflow + Invariant Proofs]
J --> K[๐ณ๏ธ 3-Model Adversarial Consensus\nDeepSeek-R1 โฅ Llama-3.3-70B โฅ Gemma-3-27B]
K --> L{2/3 Majority?}
L -->|REJECTED| E
L -->|APPROVED| M[๐ Conviction Engine\n7-Criteria Auto-Merge Gate]
M --> N[๐ Open PR + Audit Trail\nSHA-256 Tamper-Evident Log]
N --> O[๐ Training Store\nData Flywheel โ JSONL Export]
style A fill:#e94560,color:#fff
style RT fill:#d63031,color:#fff
style K fill:#6C5CE7,color:#fff
style M fill:#00b894,color:#fff
style O fill:#0984e3,color:#fff
```
## Architecture at a Glance
```mermaid
graph LR
subgraph UI["๐ฅ๏ธ Gradio Control Plane (port 7860)"]
APP[app.py\n2,311 lines]
end
subgraph ORCH["๐ง Intelligence Layer"]
HO[hermes_orchestrator.py\nAutonomous Security Research]
VL[verification_loop.py\nRetry-with-Context]
CE[conviction_engine.py\nAuto-Merge Gate]
AR[adversarial_reviewer.py\n3-Model Consensus]
end
subgraph ANALYSIS["๐ฌ Analysis Engines"]
LR2[language_runtime.py\n7 Languages]
RTF[red_team_fuzzer.py\nCEGIS Engine]
FV[formal_verifier.py\nZ3 SMT Solver]
SYM[symbolic_engine.py\nAngr Symbolic Exec]
TA[taint_analyzer.py\nDataflow Analysis]
CI[cve_intel.py\nNVD + SSEC Algorithm]
end
subgraph SECURITY["๐ก๏ธ Security Gates"]
SG[sast_gate.py\nBandit + Secrets]
SC[supply_chain.py\npip-audit + Typosquat]
VC[vuln_classifier.py\nCWE โ CVSS]
end
subgraph MEMORY["๐พ Memory & Learning"]
EM[embedding_memory.py\nSQLite / Qdrant]
TS[training_store.py\nData Flywheel]
LS[lora_scheduler.py\nFine-tune Export]
end
subgraph OUTPUT["๐ค Output Layer"]
BG[bounty_gateway.py\nHackerOne / Bugcrowd]
WH[webhook_server.py\nPort 7861]
AL[audit_logger.py\nSHA-256 Chain]
end
APP --> ORCH
APP --> ANALYSIS
APP --> SECURITY
ORCH --> MEMORY
ANALYSIS --> SECURITY
SECURITY --> OUTPUT
MEMORY --> TS --> LS
```
## Five Custom Algorithms โ Built From Scratch
VES
Vulnerability Entropy Score |
Quantifies how surprising a code path is. Combines cyclomatic complexity (via Radon), dataflow depth, and deviation from the repository's own baseline. High-VES paths are statistically anomalous execution routes that warrant deeper analysis โ the mathematical definition of "this shouldn't work this way." |
TVG
Temporal Vulnerability Graph |
A directed graph over commit history that models how a single faulty assumption propagates through the codebase as other developers build on top of it. Identifies the root-cause commit, computes blast radius, and scores the danger of downstream dependents โ giving patches a priority order. |
ACTS
Adversarial Consensus Trust Score |
Bayesian aggregation of three independent LLM adversarial reviews run concurrently. Each model votes APPROVE / REJECT / CONDITIONAL. The final score weights vote consistency, argument specificity, and historical calibration of each model against this codebase's fix patterns. Requires 2/3 majority. |
CAD
Commit Anomaly Detection |
Statistical outlier detection over git history. Computes a distribution of diff characteristics (size, churn, file types touched, message entropy) and flags commits that pattern-match against known silent security patches โ the ones developers push without saying what they really fixed. |
SSEC
Semantic Similarity Exploit Chain |
Embeds known CVE exploit patterns using microsoft/codebert-base and runs cosine similarity against repository code at the function level. Surfaces "structurally resembles CWE-X" findings even before any test failure or crash โ pure static semantic matching against 100+ historical exploit primitives. |
## The Security Research Pipeline (Hermes)
Beyond repair, the Hermes orchestrator runs a full autonomous vulnerability research sweep in six phases:
```mermaid
sequenceDiagram
participant H as ๐ง Hermes
participant R as ๐ญ RECON
participant S as ๐ฌ STATIC
participant D as ๐ฅ DYNAMIC
participant E as โ๏ธ EXPLOIT
participant C as ๐ณ๏ธ CONSENSUS
participant O as ๐ HUMAN OPERATOR
H->>R: Clone + fingerprint + map attack surface
R->>S: Attack surface map + complexity scores
S->>S: Taint analysis + CWE matching + SSEC
S->>D: Confirmed code paths + VES scores
D->>D: Generate Hypothesis PBT harnesses
D->>E: Crash payloads + stack traces
E->>E: Classify primitives: overflow/UAF/race/injection
E->>C: All findings + exploit chains
C->>C: 3-model adversarial verdict
C->>O: PENDING_HUMAN_APPROVAL
O-->>O: Human reviews + clicks Approve
O->>O: Submit to HackerOne / GitHub Advisory
```
> **Nothing is submitted to any bug bounty platform without a human clicking "Approve & Submit."** The gate is enforced at the API call level in `bounty_gateway.py` โ not just in the UI.
## Supported Languages
| Language | Detection | Test Runner | SAST Tool | Supply Chain |
|:---:|:---:|:---:|:---:|:---:|
|  | `pytest.ini` / `setup.py` | pytest / uv | Bandit + Semgrep | pip-audit |
|  | `package.json` | Jest / Mocha / Vitest | eslint-security | npm audit |
|  | `tsconfig.json` | Same as JS + tsc | Same as JS | npm audit |
|  | `pom.xml` / `build.gradle` | JUnit / TestNG | Semgrep-Java | OWASP dep-check |
|  | `go.mod` | go test | gosec | govulncheck |
|  | `Cargo.toml` | cargo test | clippy | cargo-audit |
|  | `Gemfile` | RSpec / Minitest | brakeman | bundle-audit |
Language detection is automatic. No configuration required โ Rhodawk fingerprints the cloned repository and selects the correct runtime, test runner, SAST tool, and dependency auditor.
## The MCP Server Suite โ 25 Integrated Tools
Click to expand the full MCP server manifest
| Server | Command | What It Does |
|---|---|---|
| `fetch-docs` | uvx mcp-server-fetch | Fetch CVE advisories, exploit PoCs, vendor bulletins โ 40+ security domains allowlisted |
| `github-manager` | npx @modelcontextprotocol/server-github | Create PRs, open security advisories, query commit history |
| `filesystem-research` | npx @modelcontextprotocol/server-filesystem | Read-only access to cloned repos and research scratch space |
| `memory-store` | npx @modelcontextprotocol/server-memory | Persistent knowledge graph โ exploit chains, CWE patterns, cross-session memory |
| `sequential-thinking` | npx @modelcontextprotocol/server-sequential-thinking | Structured chain-of-thought for multi-step vulnerability reasoning |
| `web-search` | npx @modelcontextprotocol/server-brave-search | Search CVEs, exploit PoCs, bug bounty writeups, research papers |
| `git-forensics` | npx @modelcontextprotocol/server-git | Deep git history: silent patches (CAD), blame tracking, anomaly detection |
| `postgres-intelligence` | npx @modelcontextprotocol/server-postgres | Query findings DB, scan history, vulnerability intelligence store |
| `sqlite-findings` | npx @modelcontextprotocol/server-sqlite | Fast queries on vulnerability metadata, CVSS scores, bounty estimates |
| `nuclei-scanner` | uvx mcp-server-shell (nuclei) | Template-based DAST, CVE detection, misconfiguration scanning |
| `semgrep-sast` | uvx mcp-server-shell (semgrep) | Taint analysis, CWE pattern matching, secrets detection โ 30+ languages |
| `trufflehog-secrets` | uvx mcp-server-shell (trufflehog) | High-signal secret scanning with 700+ detectors across git history |
| `bandit-sast` | uvx mcp-server-shell (bandit) | AST-level Python SAST: injection sinks, insecure APIs, dangerous patterns |
| `pip-audit-sca` | uvx mcp-server-shell (pip-audit) | SCA via OSV and PyPI Advisory DB โ known vulnerabilities in Python deps |
| `osv-scanner` | uvx mcp-server-shell (osv-scanner) | Multi-ecosystem SCA using the Open Source Vulnerability database (Google) |
| `z3-formal-verifier` | uvx mcp-server-shell (python3) | Z3 SMT solver โ formal verification of integer bounds and overflow invariants |
| `hypothesis-fuzzer` | uvx mcp-server-shell (hypothesis) | Property-based testing: arithmetic overflow, encoding bugs, aliasing |
| `atheris-fuzzer` | uvx mcp-server-shell (atheris) | Coverage-guided libFuzzer-backed Python fuzzing for parser bugs |
| `angr-symbolic` | uvx mcp-server-shell (python3) | angr symbolic execution โ binary analysis, path exploration, constraint solving |
| `radon-complexity` | uvx mcp-server-shell (radon) | Cyclomatic complexity + Halstead metrics + attack surface ranking |
| `ruff-linter` | uvx mcp-server-shell (ruff) | Ultra-fast linter detecting anti-patterns that correlate with security bugs |
| `aider-patcher` | uvx mcp-server-shell (aider) | Applies LLM-generated patches with diff verification and test re-run |
| `cve-intelligence` | uvx mcp-server-fetch (NVD) | Full CVE details, CVSS vectors, CWE mappings, affected version ranges |
| `bounty-platform` | uvx mcp-server-fetch | HackerOne / Bugcrowd / Intigriti / YesWeHack report submission |
| `supply-chain-monitor` | uvx mcp-server-fetch | PyPI typosquatting, dependency confusion, malicious package detection |
## The Data Flywheel
Every fix attempt โ successful or failed โ is written to a structured training store. The schema captures the complete chain:
```
failing test โ memory retrieval query โ LLM prompt โ
generated diff โ SAST results โ adversarial verdict โ test outcome โ human decision
```
This creates a proprietary fine-tuning dataset that compounds in value over time. After 50+ high-quality fixes accumulate, the LoRA scheduler exports a JSONL file ready for HuggingFace PEFT/TRL or AutoTrain:
```json
{
"messages": [
{"role": "user", "content": ""},
{"role": "assistant", "content": ""}
]
}
```
Each training cycle makes the model progressively better at fixing failures in your specific codebase. No external vendor has access to this data. It is yours.
## Required API Keys
| Variable |
Required |
Details |
GITHUB_TOKEN |
โ
Yes |
Personal Access Token with repo + security_events scopes. Used to clone repos, open PRs, and create GitHub Security Advisories. Create one here. |
OPENROUTER_API_KEY |
โ
Yes |
All LLM calls route through OpenRouter. Default models are on the free tier โ you can run this system at zero LLM cost. Get a key here. |
GITHUB_REPO |
โฌ Optional |
Target in owner/repo format. Can also be supplied at runtime via the chat UI. |
RHODAWK_AUTO_MERGE |
โฌ Optional |
Default: false. Set to true to enable autonomous PR merge when all 7 conviction criteria pass. |
RHODAWK_LORA_ENABLED |
โฌ Optional |
Default: false. Set to true to activate the LoRA fine-tune export pipeline. |
DB_BACKEND |
โฌ Optional |
Default: sqlite. Set to postgres with DATABASE_URL for production persistence. |
HACKERONE_API_KEY |
โฌ Optional |
Enables HackerOne report submission from the bounty gateway (human approval still required). |
NVD_API_KEY |
โฌ Optional |
Unlocks higher rate limits on the NIST NVD CVE API. Free to request at nvd.nist.gov. |
BRAVE_API_KEY |
โฌ Optional |
Enables Brave Search MCP tool for the Hermes web search capability. |
## Running Locally
### Step 1 โ Clone
```bash
git clone https://github.com/Rhodawk-AI/Rhodawk-devops-engine.git
cd Rhodawk-devops-engine
```
### Step 2 โ Install Python dependencies
```bash
pip install -r requirements.txt
```
> `atheris` is excluded from requirements โ it requires Clang + libFuzzer at compile time, unavailable on most CI images. The system automatically falls back to `hypothesis` for all fuzzing tasks.
### Step 3 โ Install MCP servers
```bash
npm install -g \
@modelcontextprotocol/server-github \
@modelcontextprotocol/server-memory \
@modelcontextprotocol/server-filesystem \
@modelcontextprotocol/server-sequential-thinking \
@modelcontextprotocol/server-brave-search \
@modelcontextprotocol/server-git
```
### Step 4 โ Configure environment
```bash
export GITHUB_TOKEN="ghp_your_token_here"
export OPENROUTER_API_KEY="sk-or-your_key_here"
export GITHUB_REPO="owner/repo" # optional โ can set in UI
mkdir -p /data
```
### Step 5 โ Run
```bash
python -u app.py
```
Gradio UI: `http://localhost:7860`
Webhook server: `http://localhost:7861`
## Docker
```bash
# Build
docker build -t rhodawk-ai .
# Run
docker run -d \
-p 7860:7860 \
-p 7861:7861 \
-v rhodawk_data:/data \
-e GITHUB_TOKEN="ghp_your_token_here" \
-e OPENROUTER_API_KEY="sk-or-your_key_here" \
-e GITHUB_REPO="owner/target-repo" \
rhodawk-ai
```
## HuggingFace Spaces Deployment
```
1. Go to: huggingface.co/spaces/Architect8999/rhodawk-ai-devops-engine
2. Duplicate the Space (top-right button)
3. Add Secrets in Space Settings:
GITHUB_TOKEN โ your GitHub PAT
OPENROUTER_API_KEY โ your OpenRouter key
4. The Space builds and runs automatically via the included Dockerfile
```
## Event-Driven Mode โ GitHub Webhook
Make Rhodawk trigger automatically on every CI failure:
```
GitHub repo โ Settings โ Webhooks โ Add webhook
Payload URL: https://your-space.hf.space/webhook/github
Content type: application/json
Secret: (set RHODAWK_WEBHOOK_SECRET to the same value)
Events: Push, Check runs, Status
```
From this point forward, every failing CI run triggers the full autonomous repair loop with no manual intervention.
Supported webhook endpoints:
```
POST /webhook/github GitHub push / check_run / status (HMAC-SHA256 validated)
POST /webhook/ci Generic CI failure payload (any CI system)
POST /webhook/trigger Manual trigger with repo + test path
GET /webhook/health Health check
GET /webhook/queue Current job queue status
```
## Repository Structure
Click to expand โ all 42 source files with descriptions
```
rhodawk-devops-engine/
โ
โโโ ๐๏ธ CONTROL PLANE
โ โโโ app.py Main entry point. Gradio UI + full audit loop. (2,311 lines)
โ โโโ webhook_server.py Event-driven server on port 7861. GitHub/CI webhooks.
โ
โโโ ๐ง INTELLIGENCE
โ โโโ hermes_orchestrator.py 6-phase autonomous security research agent. (715 lines)
โ โโโ adversarial_reviewer.py 3-model concurrent consensus code review.
โ โโโ verification_loop.py Retry-with-context fix loop.
โ โโโ conviction_engine.py 7-criteria auto-merge gate.
โ
โโโ ๐ LANGUAGE RUNTIMES
โ โโโ language_runtime.py Python/JS/TS/Java/Go/Rust/Ruby abstraction. (1,540 lines)
โ
โโโ ๐ด RED TEAM ENGINE
โ โโโ red_team_fuzzer.py CEGIS autonomous attack engine. (1,561 lines)
โ
โโโ ๐ฌ ANALYSIS
โ โโโ taint_analyzer.py Dataflow taint: source-to-sink tracking.
โ โโโ symbolic_engine.py Angr symbolic execution + path exploration.
โ โโโ formal_verifier.py Z3 SMT: integer overflow + invariant proofs.
โ โโโ fuzzing_engine.py Hypothesis PBT harness generator.
โ โโโ exploit_primitives.py Overflow / UAF / race / injection classification.
โ โโโ harness_factory.py PoC harness compiler for operator-reviewed gaps.
โ โโโ chain_analyzer.py Multi-primitive vulnerability chain synthesizer.
โ โโโ commit_watcher.py CAD: silent security patch detection.
โ โโโ repo_harvester.py Autonomous target repository selection.
โ โโโ semantic_extractor.py AST-level feature extraction for VES scoring.
โ
โโโ ๐ก๏ธ SECURITY GATES
โ โโโ sast_gate.py Bandit + 16-pattern secret scanner.
โ โโโ supply_chain.py pip-audit + typosquatting detection.
โ โโโ vuln_classifier.py CWE taxonomy โ CVSS scoring โ severity.
โ โโโ cve_intel.py NVD/CVE API + SSEC algorithm.
โ
โโโ ๐พ MEMORY & LEARNING
โ โโโ embedding_memory.py Dual-backend: SQLite/MiniLM or Qdrant/CodeBERT.
โ โโโ memory_engine.py Fix outcome tracking + similarity retrieval.
โ โโโ training_store.py SQLite/Postgres training data flywheel.
โ โโโ lora_scheduler.py LoRA fine-tune export scheduler.
โ
โโโ ๐ค OUTPUT & DISCLOSURE
โ โโโ bounty_gateway.py HackerOne / Bugcrowd / GitHub Advisories gateway.
โ โโโ disclosure_vault.py 90-day coordinated disclosure timeline vault.
โ โโโ audit_logger.py Append-only SHA-256 tamper-evident audit trail.
โ โโโ public_leaderboard.py Fix success rate leaderboard.
โ
โโโ โ๏ธ INFRASTRUCTURE
โ โโโ github_app.py GitHub App JWT authentication.
โ โโโ job_queue.py Job queue with status tracking + metrics.
โ โโโ worker_pool.py Parallel audit worker pool.
โ โโโ notifier.py Slack/webhook notification dispatch.
โ โโโ swebench_harness.py SWE-bench Verified evaluation harness.
โ
โโโ ๐ฆ CONFIGURATION
โ โโโ mcp_config.json 25-server MCP suite configuration (template, no secrets).
โ โโโ Dockerfile Two-stage build: Python 3.12-slim + Node.js for MCP.
โ โโโ requirements.txt Python dependencies (31 packages).
โ โโโ FOUNDER_PLAYBOOK.md Full technical + investor documentation. (1,119 lines)
โ โโโ SECURITY_RESEARCH_PLAYBOOK.md Ethical AVR operator guide.
```
## Security by Design
| Principle | Implementation |
|---|---|
| **No hardcoded secrets** | Every credential is loaded from environment variables. The codebase contains zero API keys. |
| **MCP runtime injection** | `mcp_config.json` is a template. Secrets are written to `/tmp/mcp_runtime.json` at startup โ never committed. |
| **Tamper-evident audit trail** | `audit_logger.py` maintains a SHA-256 chain across all log entries. Any modification to historical records is detectable. |
| **Human-gated disclosure** | `bounty_gateway.py` enforces approval at the API call level. Removing the UI button does not bypass the gate. |
| **Formal patch verification** | Z3 proves bounded integer invariants on every AI-generated diff before any merge can occur. |
| **SSRF prevention** | All MCP fetch tools operate against `FETCH_ALLOWED_DOMAINS` allowlists. Outbound requests are restricted to explicitly permitted security domains. |
| **Coordinated disclosure** | 90-day Google Project Zero-standard disclosure timeline tracked per finding in `disclosure_vault.py`. |
## Default LLM Models
All default models are on OpenRouter's free tier. This system runs at zero LLM cost out of the box.
| Role | Default Model | Override Variable |
|---|---|---|
| Code Fix Generation | `qwen/qwen-2.5-coder-32b-instruct:free` | `RHODAWK_MODEL` |
| Hermes Orchestrator | `deepseek/deepseek-r1:free` | `HERMES_MODEL` |
| Hermes Fast Tasks | `deepseek/deepseek-v3:free` | `HERMES_FAST_MODEL` |
| Adversarial Review #1 | `deepseek/deepseek-r1:free` | `RHODAWK_ADVERSARY_MODEL` |
| Adversarial Review #2 | `meta-llama/llama-3.3-70b-instruct:free` | hardcoded fallback |
| Adversarial Review #3 | `google/gemma-3-27b-it:free` | hardcoded fallback |
**Every feature in this README is implemented in the files above.**
**No mocks. No stubs. No vaporware. The pipeline runs end-to-end.**
[](https://huggingface.co/spaces/Architect8999/rhodawk-ai-devops-engine)
Rhodawk AI ยท Autonomous DevSecOps Control Plane v4.0 ยท Proprietary License