Nano-analyzer

A minimal LLM-powered zero-day vulnerability scanner by AISLE.

Research prototype for demonstration purposes. This is a simple, single-file harness that is able to detect real zero-day vulnerabilities. Note that it is a prototype, biased towards C/C++ memory safety bugs, and will produce false positives. We are sharing it as-is in the spirit of open research — expect sharp corners.

What it does

Nano-analyzer is a simple single-file Python scanner that sends source code through a three-stage LLM pipeline:

1. Context generation — a model writes a security briefing about the file: what it does, where untrusted data flows, which buffers exist and how big they are.
2. Vulnerability scan — the same model, primed with the context, hunts for zero-day bugs function by function and outputs structured findings.
3. Skeptical triage — each finding is challenged over multiple rounds by a skeptical reviewer that can grep the codebase to verify (or refute) defenses. An arbiter makes the final call.

Results are saved as Markdown and JSON files for human review.

Current limitations

This is a v0.1 prototype. Please keep the following in mind:

• C/C++ bias. The prompts, few-shot examples, and heuristics are heavily tuned for C/C++ memory safety vulnerabilities (buffer overflows, NULL derefs, integer overflows, type confusion). It will scan other languages but is much less effective there.
• False positives. Even with multi-round triage, expect findings that don't hold up on closer inspection. Always verify manually.
• False negatives. The scanner can miss entire vulnerability classes — logic bugs, race conditions, cryptographic issues, authentication bypasses, etc. A clean scan does not mean the code is safe.
• Single-file analysis. Each file is scanned independently. Cross-file vulnerabilities that depend on interactions between compilation units will likely be missed.
• LLM-dependent. Results vary with the model used. Different models will find different things and hallucinate different false positives.

Setup

Requirements

• Python 3.8+
• An OpenAI API key (for OpenAI models) or an OpenRouter API key (for other providers)
• Optional: ripgrep (rg) for triage grep lookups
• Optional: Google codesearch (csearch/cindex) for faster grep on large repos

Install

git clone https://github.com/weareaisle/nano-analyzer.git
cd nano-analyzer
# No dependency installation needed. Run directly:
python3 scan.py --help

API keys

Set your API key as an environment variable:

# For OpenAI models (model names without a slash, e.g. "gpt-5.4-nano"):
export OPENAI_API_KEY=sk-...

# For OpenRouter models (model names with a slash, e.g. "qwen/qwen3-32b"):
export OPENROUTER_API_KEY=sk-or-...

The scanner determines which key to use based on the model name: if it contains a /, it routes through OpenRouter; otherwise it uses the OpenAI API directly.

Usage

Basic scan

# Scan a single file
python3 scan.py ./path/to/file.c

# Scan a directory recursively
python3 scan.py ./path/to/src/

Common options

# Use a different model
python3 scan.py ./src --model gpt-5.4

# Control parallelism
python3 scan.py ./src --parallel 30

# Point triage grep at the full repo root (useful when scanning a subdirectory)
python3 scan.py ./lib/crypto/ --repo-dir ./

# Only surface high-confidence findings
python3 scan.py ./src --min-confidence 0.7

# More triage rounds for higher accuracy (default: 5)
python3 scan.py ./src --triage-rounds 7

All flags

Flag	Default	Description
path	(required)	File or directory to scan
--model	gpt-5.4-nano	Model for all stages (context, scan, triage)
--parallel	50	Max concurrent scan API calls
--triage-threshold	medium	Triage findings at or above this severity
--triage-rounds	5	Triage rounds per finding
--triage-parallel	50	Max concurrent triage API calls
--max-connections	parallel + triage-parallel	Total API call cap
--min-confidence	0.0	Only show findings above this confidence (0.0–1.0)
--project	directory name	Project name used in triage prompts
--repo-dir	auto	Repo root for grep lookups (auto: parent dir for files, scan dir for folders)
--output-dir	~/nano-analyzer-results/<timestamp>/	Where to save results
--max-chars	200,000	Skip files larger than this
--verbose-triage	off	Show per-round triage progress

Output

Results are saved to ~/nano-analyzer-results/<timestamp>/ (or --output-dir):

<timestamp>/
├── summary.json              # machine-readable scan summary
├── summary.md                # human-readable scan summary
├── <filename>.md             # raw scanner output per file
├── <filename>.context.md     # context briefing per file
├── <filename>.json           # full result data per file
├── triages/                  # detailed triage reasoning
│   └── T0001_<file>_<title>.md
├── findings/                 # findings that survived triage
│   └── VULN-001_<file>.md
├── triage.json               # all triage verdicts
└── triage_survivors.md       # summary of validated findings

How triage works

When a scan finds a medium-or-above severity issue, the triage pipeline kicks in:

1. A skeptical reviewer examines the finding against the actual code and can grep the codebase to verify or refute claimed defenses.
2. This repeats for multiple rounds (default: 5), with each reviewer seeing prior arguments and encouraged to find new evidence rather than rehash old points.
3. A final arbiter reads all rounds and makes a VALID/INVALID call.
4. The confidence score (e.g. 80% [VVIVV→V]) reflects the fraction of rounds that said VALID.

Findings that survive triage are written to the findings/ directory with full reasoning chains.

Disclaimer

This tool is a research prototype. It is not a replacement for professional security audits, manual code review, or established static analysis tools. Do not rely on it as your sole security assessment. Use at your own risk.

License

Apache License 2.0