The AI security engineer that proves its fixes.
Paste a GitHub repo → find what's exploitable → fix it → prove the fix → plain-English report.
Built at Tokyo AgentForge 2026 with Daytona · Nosana · Tavily
🌐 Live showcase · 📑 Pitch deck (PDF) · 📊 Results · 🚀 Run it
Most breaches start with a known bug. Teams don't patch because they can't tell what's actually exploitable, and upgrades break code. AI review bots comment; PatchPilot executes every fix and proves it.
| Repository | Stack | Verdict | Key finding | Time |
|---|---|---|---|---|
| Agent-Field/CodeAF | Go | 🟠 | 22 known vulns across 110 modules (x/net, x/crypto, x/image) |
81 s |
| digininja/DVWA | PHP | 🔴 | 9 critical: command injection, SQL injection, file inclusion, all with fixes | 61 s |
| OWASP/NodeGoat | Node | 🔴 | eval() remote code execution + 177 known vulns |
91 s |
| microsoft/generative-ai-for-beginners | Python | 🟢 | Clean: 66 deps, no known vulns | 106 s |
demo_repo |
Python | ✅ | Auto-patched 20 → 0, fixed the Jinja2 break, tests 3/3 | ~110 s |
Cost per scan: $0.002–0.013 · Full reports in cases/
| 🔴 DVWA (PHP) | 🟠 CodeAF (Go) |
|---|---|
![]() |
![]() |
| 🔴 NodeGoat (Node) | 🟢 Microsoft repo |
![]() |
![]() |
| Step | What happens | Powered by |
|---|---|---|
| ① Sandbox | Disposable VM (boots ~1 s); untrusted code never touches your laptop | Daytona |
| ② Recon | Reads the repo's architecture and attack surface first | Tavily → DeepWiki |
| ③ Hunt | Two parallel sub-agents: Threat Intel (OSV · EPSS · CISA KEV · news) + Code Review | OSV · FIRST · CISA · Tavily |
| ④ Fix | Agent loop: think → run a tool in the sandbox → observe → repeat | Nosana (Qwen) |
| ⑤ Prove | Re-install, re-scan, re-test → diff, PR text, report | Daytona |
- Daytona:
create·process.exec·git.clone·fs·delete; 4 repos scanned in 4 sandboxes at once. - Nosana: OpenAI-compatible gateway (Qwen 3.8 27B, tool calling) + our own RTX 3080 GPU (Qwen 3.5 9B, $0.10/h) with auto-failover and live credit tracking.
- Tavily:
extract(DeepWiki recon) ·searchwith answers (fix research) · news/forum search (threat intel).
| Daytona sandboxes | Nosana GPU | Tavily usage |
|---|---|---|
![]() |
![]() |
![]() |
- Isolated: every command the AI chooses runs in a throwaway Daytona VM that is deleted after each scan.
- Verified: the Prove step ignores the agent's claims and re-scans and re-tests independently.
- Transparent: you watch every thought, tool call and sandbox command live; reports cite their sources.
- Auto-fix ships for Python today; PHP/Node/Go are report-only (with fixes suggested).
- Next: AgentField CodeAF as the in-sandbox fixer (
codeaf do, pointed at Nosana), a GitHub App that opens verified PRs, and a CI gate.
git clone https://github.com/Ana2k/patch-pilot && cd patch-pilot
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
cp .env.example .env # DAYTONA_API_KEY, NOSANA_API_KEY, TAVILY_API_KEY
.venv/bin/uvicorn server:app --port 8000Open http://localhost:8000, paste any public repo (or leave it empty for the demo) and click Run security test. You'll watch the 5 stages light up, the agent's thoughts, and the sandbox terminal live; the report lands in ~1–2 min.
Saved reports open instantly: /#case=CodeAF · /#case=DVWA · /#case=NodeGoat.
agent.py orchestrator + fixer · security.py sub-agents + report · scanners.py PHP/Go/JS/Python checks · server.py API · static/ site · demo_repo/ fake secrets only. Scanned code runs only inside Daytona.






