Clearwing

By Eric Hartford, Lazarus AI

Inspired by Anthropic's Glasswing.

The challenge: Produce similar results as Glasswing - using models everyone has access to.

Autonomous vulnerability scanner and source-code hunter. Built on genai-pyo3, a native Rust-backed LLM runtime speaking every major provider (Anthropic, OpenAI, OpenRouter, Ollama, LM Studio, Together, Groq, DeepSeek, MiniMax, Gemini, any OpenAI-compatible endpoint).

Clearwing is a dual-mode offensive-security tool:

• Network-pentest agent — a ReAct-loop agent with 63 bind-tools that scans live targets, detects services and vulnerabilities, runs sandboxed Kali tools, attempts exploits (gated through a human-approval guardrail), and writes reports to a persistent knowledge graph.
• Source-code hunter — a file-parallel agent-driven pipeline that ranks source files, fans out per-file hunter agents (full-shell or constrained), uses ASan/UBSan crashes as ground truth, verifies findings with a 4-axis validator (REAL / TRIGGERABLE / IMPACTFUL / GENERAL), runs PoC stability checks across fresh containers, optionally generates validated patches, and emits SARIF/markdown/JSON reports with explicit evidence levels (suspicion → static_corroboration → crash_reproduced → root_cause_explained → exploit_demonstrated → patch_validated). Features three-band budget promotion, entry-point sharding for large files, cross-subsystem hunting, a shared findings pool with root-cause deduplication, multi-turn agentic exploit development, and human-in-the-loop exploit elaboration.
• N-day exploit pipeline — given CVE IDs, builds the vulnerable version, develops working exploits, and validates against the patched version to confirm the fix.
• Reverse engineering pipeline — decompiles closed-source ELF binaries via Ghidra, reconstructs plausible source with an LLM, then hunts vulnerabilities using a hybrid source + binary validation approach.
• Campaign orchestration — runs sourcehunt across dozens or hundreds of repositories from a single YAML config with shared budget, checkpoint/resume, and aggregate reporting.
• Responsible disclosure — human-in-the-loop validation workflow with MITRE/HackerOne template generation, SHA-3 cryptographic commitments for provable priority, timeline tracking, and batched disclosure.
• Benchmarking & evaluation — OSS-Fuzz crash severity ladder for model comparison, and an A/B testing framework for measuring whether preprocessing helps or hurts finding quality.

Authorized use only. Clearwing is a dual-use offensive-security tool. Run it only against targets you own or have explicit written authorization to test. Operators are responsible for scope, legal authorization, and disclosure. See SECURITY.md.

Install

End users — install the tagged release straight from GitHub:

git clone https://github.com/Lazarus-AI/clearwing.git
cd clearwing

# uv sync is recommended because Clearwing pins genai-pyo3 through
# tool.uv.sources in pyproject.toml.
uv sync --all-extras
source .venv/bin/activate  # fish: source .venv/bin/activate.fish

# Interactive setup wizard — menu-driven provider selection,
# credential entry, optional live test, persists to ~/.clearwing/config.yaml
clearwing setup

# Environment check — verifies Python, credentials, Docker daemon,
# external tools, optional extras, and network reachability
clearwing doctor

clearwing --version   # 1.0.0
clearwing --help

Or skip the wizard and configure directly:

# Anthropic direct
export ANTHROPIC_API_KEY=sk-ant-...

# Or any OpenAI-compatible endpoint — OpenRouter, Ollama, LM Studio,
# vLLM, Together, Groq, DeepSeek, OpenAI:
export CLEARWING_BASE_URL=https://openrouter.ai/api/v1
export CLEARWING_API_KEY=sk-or-...
export CLEARWING_MODEL=anthropic/claude-opus-4

See docs/providers.md for provider-specific recipes and per-task routing.

Developers — clone and install the locked development environment:

git clone https://github.com/Lazarus-AI/clearwing.git
cd clearwing
uv sync --all-extras
source .venv/bin/activate  # fish: source .venv/bin/activate.fish
clearwing --help

Requirements: Python 3.10+ and optionally Docker for the Kali container and sanitizer-image sandbox features. genai-pyo3 ships as prebuilt wheels on PyPI (linux x86_64/aarch64, macOS universal2, windows x86_64, Python 3.9–3.13), so no Rust toolchain is needed for installation.

Quickstart

# Network scan a single target
clearwing scan 192.168.1.10 -p 22,80,443 --detect-services

# Source-code hunt a repo (standard depth — sandboxed LLM hunters,
# adversarial verifier, mechanism memory, variant loop)
clearwing sourcehunt https://github.com/example/project \
    --depth standard

# Proof-carrying hunt — typed facts, obligations, evidence, falsification,
# and auditable finding/rejection/incomplete certificates
clearwing sourcehunt /path/to/project \
    --flow proof \
    --compile-commands compile_commands.json \
    --budget 50

# N-day exploit pipeline — build and exploit known CVEs
clearwing sourcehunt https://github.com/example/project \
    --nday --cve-list CVE-2024-1234,CVE-2024-5678

# Reverse engineering — hunt vulnerabilities in closed-source binaries
clearwing sourcehunt /path/to/binary --reveng --arch x86_64

# Campaign-scale orchestration across multiple projects
clearwing campaign run campaign.yaml

# Responsible disclosure workflow
clearwing disclose queue ./results/sourcehunt/sh-*/
clearwing disclose review

# OSS-Fuzz crash severity benchmark
clearwing bench ossfuzz --corpus-dir ./oss-fuzz-projects --mode standard

# A/B test whether preprocessing helps or hurts
clearwing eval preprocessing --project https://github.com/example/project \
    --configs glasswing_minimal,sourcehunt_full --runs 3

# Interactive ReAct chat with the full tool set
clearwing interactive

# Non-interactive CI mode with SARIF output for GitHub Code Scanning
clearwing ci --config .clearwing.ci.yaml --sarif results.sarif

See docs/quickstart.md for a fuller walkthrough including credentials, session resume, and mission-mode operation.

Proof-carrying sourcehunt (--flow proof)

clearwing sourcehunt has two separate investigation engines. The default, --flow legacy, uses the file-oriented hunter, verifier, and exploiter pipeline. The opt-in --flow proof engine moves investigation state out of the model context and into typed, inspectable artifacts:

repository snapshot
  → extracted facts and completeness manifest
  → invariant-oriented candidates and threat models
  → proof-plan obligation graphs
  → bounded mechanical, model, and dynamic actions
  → independent falsification
  → finding, rejection, or incomplete certificates

The proof engine uses models for bounded judgments rather than asking one agent to remember the whole investigation. Its default local-first policy tries the configured proof_local route first and escalates an unresolved atomic question to proof_frontier. Exploratory work receives a separate, bounded share of the action budget and cannot become a reportable finding without satisfying the same evidence obligations.

For a C or C++ repository, generate compile_commands.json from the real build and run Docker before starting. The path passed to --compile-commands is resolved relative to the target repository. Proof preflight fails closed when the compilation database or analysis sandbox is unavailable; it does not silently fall back to lexical extraction or the legacy engine.

clearwing sourcehunt /path/to/project \
  --flow proof \
  --compile-commands compile_commands.json \
  --model-routing local-first \
  --structured-budget 90% \
  --exploration-budget 10% \
  --proof-max-actions 200 \
  --proof-max-model-calls 40 \
  --proof-max-dynamic-actions 20 \
  --retain-incomplete-certificates \
  --emit-rejection-certificates \
  --falsify \
  --budget 50 \
  --output-dir ./results/sourcehunt-proof

--budget is the run-wide monetary cap; omit it or pass --budget 0 for the unlimited default. The three --proof-max-* flags separately bound all proof actions, model calls, and dynamic actions. Dynamic validation is performed only from a typed --validation-manifest inside the sandbox; add --gvisor when that runtime is installed for an additional isolation layer.

Each session is written under <output-dir>/sh-<session-id>/. Start with manifest.json, which records the engine, snapshot, blind boundary, run status, spend, action counts, certificate counts, and output index. The session also retains append-only facts, candidates, obligations, actions, claims, evidence, derivations, context packets, proof graphs, falsification results, and content-addressed runtime artifacts. Human and tool consumers can use report.md, findings.json, and findings.sarif; only accepted finding certificates appear in JSON and SARIF. Rejected and unresolved work remains under certificates/rejections/ and certificates/incomplete/.

An exit status of 3 means the proof is incomplete or the run-wide budget was exhausted. It is a preserved partial investigation, not proof that the repository is safe and not necessarily a process failure.

See the FFmpeg proof-flow walkthrough for an end-to-end blind C/C++ example, the design document for the evidence model, and the evaluation and rollout guide before comparing or promoting proof-flow results.

Running sourcehunt on a local repo (FFmpeg example)

The positional repository argument accepts either a Git URL or a local checkout. A local checkout is preferable for a reproducible FFmpeg proof run because it can be pinned to an exact commit and built before discovery:

git clone https://code.ffmpeg.org/FFmpeg/FFmpeg.git
cd FFmpeg

./configure \
  --cc=clang \
  --cxx=clang++ \
  --disable-doc \
  --disable-stripping \
  --enable-debug=3

bear -- make -j"$(getconf _NPROCESSORS_ONLN 2>/dev/null || sysctl -n hw.ncpu)"

clearwing sourcehunt "$PWD" \
  --flow proof \
  --compile-commands compile_commands.json \
  --model-routing local-first \
  --falsify \
  --budget 50 \
  --output-dir ../results/sourcehunt-ffmpeg-proof

This is only the shortest useful invocation. The full walkthrough pins the known vulnerable parent, preserves a strict blind boundary, captures an ASan/UBSan build, separates local and frontier model routes, validates with a typed runtime manifest, and explains every proof artifact. Follow docs/FFmpeg.md when reproducing the case or collecting evaluation results. See docs/providers.md for managed and self-hosted provider configuration.

Architecture at a glance

┌──────────────────────┐      ┌────────────────────────────────┐
│ Network-pentest agent│      │ Source-code hunter             │
│ clearwing.agent.graph│      │ clearwing.sourcehunt.runner    │
│  (63 tools, ReAct)   │      │                                │
│                      │      │ preprocess → rank → pool →     │
│                      │      │   hunter → verify → exploit →  │
│                      │      │   variant loop → auto-patch →  │
│                      │      │   report                       │
└─────────┬────────────┘      └────────┬───────────────────────┘
          │                             │
          └───────────┬─────────────────┘
                      ▼
┌───────────────────────────────────────────────────────────────┐
│ N-day pipeline │ Reveng pipeline │ Campaign orchestrator      │
│ Disclosure workflow + SHA-3 commitments                       │
├───────────────────────────────────────────────────────────────┤
│                    Shared substrate                          │
│  Finding dataclass  │  capabilities probe  │  sandbox layer  │
│  knowledge graph    │  episodic memory     │  event bus      │
│  telemetry          │  guardrails + audit  │  CVSS scoring   │
│  artifact store     │  behavior monitor    │  seccomp        │
├───────────────────────────────────────────────────────────────┤
│  Bench: OSS-Fuzz severity ladder  │  Eval: preprocessing A/B │
└───────────────────────────────────────────────────────────────┘

Deep dives live in docs/:

Doc	What it covers
docs/index.md	Landing page + table of contents
docs/quickstart.md	Full install + first run walkthrough
docs/providers.md	OpenRouter / Ollama / LM Studio / vLLM / Together / Groq recipes, per-task routing, env-var precedence
docs/architecture.md	Both pipelines, substrate, capability gating, tool layout
docs/cli.md	Every subcommand flag, grouped by workflow
docs/api.md	API reference (mkdocstrings autogen)

Once the GitHub Pages workflow ships, docs will be hosted at https://lazarus-ai.github.io/clearwing/.

Development

uv sync --all-extras
source .venv/bin/activate  # fish: source .venv/bin/activate.fish
pytest -q
ruff check clearwing/ tests/
ruff format --check clearwing/ tests/
mypy --follow-imports=silent \
  clearwing/findings \
  clearwing/sourcehunt \
  clearwing/capabilities.py \
  clearwing/agent/tools \
  clearwing/core
python -m mkdocs serve --dev-addr 127.0.0.1:8000

See CONTRIBUTING.md for the full dev-setup guide and PR checklist.

Reporting vulnerabilities

There are two lanes, and they go to different places:

• Vulnerabilities in Clearwing → GitHub Security Advisories (https://github.com/Lazarus-AI/clearwing/security/advisories/new). See SECURITY.md for scope, SLA, and safe-harbor.
• Vulnerabilities Clearwing finds in someone else's software → that vendor's disclosure channel. clearwing sourcehunt --export-disclosures generates pre-filled MITRE CVE-request and HackerOne templates for every finding at evidence_level >= root_cause_explained.

License

MIT. See LICENSE.