🛡️ PatchPilot v3

Find what's exploitable. Fix it. Prove the fix.

PatchPilot is an AI security engineer. Paste a GitHub repository and it opens the code in a throwaway sandbox, hunts for vulnerabilities, ranks them by real-world attack risk, patches what it safely can, and re-runs the tests to prove nothing broke.

Try: · ·
1

Sandbox

Daytona
A disposable computer in the cloud. Untrusted code runs here, never on yours.
ready
Isolate the danger

Every command the AI chooses runs inside a fresh Daytona VM that is deleted afterwards. Daytona boots in under a second, so each scan gets a clean machine.

sb = Daytona(cfg).create(); sb.process.exec("pytest -q")
–boot time
–vCPU · RAM · disk
2

Recon

Tavily + DeepWiki
Reads the project's docs to learn what it does and where attackers get in.
ready
Understand before touching

Tavily extracts DeepWiki's AI-generated docs for the repo as clean text, so the agent maps the attack surface the way a senior engineer reads the architecture first.

tav.extract(urls=["https://deepwiki.com/owner/repo"])
0Tavily calls
–recon source
3

Hunt

OSV · SAST · EPSS · KEV
Two sub-agents work in parallel: one checks dependencies, one reviews the code.
Threat intelCode review
ready
Rank by real attack risk

Known flaws come from OSV.dev. FIRST EPSS predicts the chance of attack in the next 30 days, and CISA KEV flags flaws already exploited in the wild. Static analysis runs inside the sandbox and the LLM removes false positives.

GET api.first.org/data/v1/epss?cve=CVE-2023-32681
–dependency advisories
–code findings
4

Fix

Nosana GPU · Qwen
An open AI model upgrades packages and repairs code the upgrade breaks.
ready
An agent is a loop

The model runs on Nosana's decentralized GPUs. It picks one tool per step (run, read, write, search), sees the result, and repeats until the tests pass. It's cheap enough to run on every pull request.

llm.chat.completions.create(model=m, tools=TOOLS)
–model
$0.000this run
0tokens in · out
–Nosana credits left
5

Prove

Daytona
Re-installs, re-scans and re-runs the tests. No proof, no patch.
ready
Never trust "done"

PatchPilot ignores the AI's claim and checks for itself: a fresh vulnerability scan plus the project's own test suite, in the same sandbox, before writing the report.

before 20 advisories → after 0 · tests 3/3 ✓
–before → after
–tests
Report

🧠 What the agent is thinking

🖥️ Sandbox terminal runs remotely in Daytona

Case studies

Real scans of public repositories. Open one to read the full report, no waiting.

Loading…

How each partner powers it

Each one does a job the others can't.

Daytona

isolation

A fresh VM for each scan. Clones the repo, runs SAST, installs and tests, then is deleted. Boots in under a second.

sandbox.git.clone(url, path)

Nosana

the brain

Open Qwen model on decentralized GPUs, with tool calling. Falls back automatically from the shared gateway to our own RTX 3080 deployment.

base_url=inference.nosana.com/v1

Tavily

fresh knowledge

Migration guides, error fixes and security news the model's frozen weights don't know. Also pulls DeepWiki pages as clean text.

tav.search(q, topic="news")

OSV · CISA · FIRST

ground truth

OSV lists known flaws, CISA KEV says which are exploited today, and FIRST EPSS predicts which will be next.

api.osv.dev/v1/query

How an AI agent and a sandbox work together

It's a loopThe AI thinks, takes one action, looks at the result, and repeats. That loop is the agent.
Tools, not magicThe AI only picks an action. Our code runs it and hands back the output.
A safe roomActions happen in a disposable sandbox, so a bad command can't hurt anything real.
Proof over promisesResults are re-checked independently before anything reaches you.