sectalon 0.1.2

Talon CLI: phishing URL analysis with Playwright, Ollama/OpenAI, and evidence capture.

🦅 Talon (Alpha)

The Autonomous AI Phishing Hunter Detonate suspicious URLs, capture evidence, and get a fast phishing-risk verdict.

Talon is a practical phishing URL analysis tool that:

• detonates suspicious links in headless Chromium,
• captures forensic evidence (screenshot + DOM snapshot),
• returns an LLM-assisted phishing risk verdict (with heuristic fallback).

---

✨ Features (V1)

• 🕵️ URL detonation: opens a target URL safely in Playwright.
• 🔁 Redirect awareness: records the final resolved URL after redirects.
• 🧾 Evidence capture: stores full-page screenshot and DOM snapshot.
• 🤖 LLM analysis: uses an LLM for smarter risk reasoning.
• 📊 Structured output: writes a JSON report with score, level, reasons, and method.
• 🧱 Docker sandbox mode: runs analysis in a hardened container profile.

---

⚙️ How It Works

1. Ingest: receive a URL input.
2. Detonate: load it in headless Chromium.
3. Collect: save final URL, HTTP status, title, screenshot, and DOM.
4. Verdict: run LLM analysis (ollama / openai) with heuristic fallback.

---

📁 Project Structure

.
├── talon_v1.py
├── requirements.txt
├── Dockerfile
├── docker-compose.sandbox.yml
├── .gitignore
└── evidence/              # generated at runtime (ignored by git)

---

🚀 Local Setup

git clone https://github.com/tejkdno1/Talon.git
cd Talon
python3 -m pip install -r requirements.txt
python3 -m playwright install chromium

---

📥 Install As CLI Package (sectalon)

You can install Talon on another machine as a CLI package and run it via sectalon.

From GitHub:

python3 -m pip install "git+https://github.com/tejkdno1/Talon.git"

Or with pipx (recommended for CLI tools):

pipx install "git+https://github.com/tejkdno1/Talon.git"

Then run:

sectalon "https://example.com" --llm-provider ollama --llm-model gemma4

Notes:

• Package name is sectalon, command name is sectalon.
• First run auto-downloads Chromium if missing.

---

🧠 Onboarding + Intent Mode

Run interactive onboarding (OpenClaw-style) to connect your LLM backend:

sectalon onboard

This stores config at:

~/.config/sectalon/config.json

Then you can parse natural-language intent through your connected LLM:

sectalon intent "check if amezon.in looks like phishing"

Supported intent outputs include:

• analyze_url
• onboard
• help
• unknown

---

📦 One-File Executable (Open-Source Friendly)

Build a single binary (no Python required on target machine):

./build_onefile.sh

Output:

dist/sectalon

Run it directly:

./dist/sectalon "https://example.com" --llm-provider ollama --llm-model gemma4

Notes:

• Keep Ollama running locally when using --llm-provider ollama.
• For OpenAI provider, set OPENAI_API_KEY as usual.
• On first run, Chromium auto-installs to ~/.cache/ms-playwright.

---

▶️ Quick Start (Host Run)

python3 talon_v1.py "https://example.com"

Local Ollama (Gemma4) - recommended for your setup

Start Ollama and pull model:

ollama pull gemma4

Run Talon using Ollama backend:

export TALON_LLM_PROVIDER="ollama"
export TALON_LLM_MODEL="gemma4"
export OLLAMA_HOST="http://localhost:11434"
export OLLAMA_TIMEOUT_SEC="180"
python3 talon_v1.py "https://example.com"

OpenAI (optional)

If you want cloud LLM instead:

export OPENAI_API_KEY="your_api_key_here"
export TALON_LLM_PROVIDER="openai"
python3 talon_v1.py "https://example.com"

Optional:

python3 talon_v1.py "example.com/login" --output-dir evidence --timeout-ms 20000

Force heuristic-only mode:

python3 talon_v1.py "https://example.com" --no-llm

Optional model override:

export TALON_LLM_MODEL="gemma4"

Optional provider override per run:

python3 talon_v1.py "https://example.com" --llm-provider ollama --llm-model gemma4

---

🛡️ Docker Sandbox Run (Recommended)

Build once:

docker compose -f docker-compose.sandbox.yml build

Run analysis:

TARGET_URL="https://leadscruise.com" docker compose -f docker-compose.sandbox.yml run --rm talon

For Docker + host Ollama, default OLLAMA_HOST is set to:

http://host.docker.internal:11434

You can override it if needed:

OLLAMA_HOST="http://host.docker.internal:11434" TARGET_URL="https://example.com" docker compose -f docker-compose.sandbox.yml run --rm talon

The sandbox profile includes:

• read-only root filesystem,
• all Linux capabilities dropped,
• no-new-privileges,
• CPU/memory/PID limits,
• output only through mounted ./evidence.

---

📦 Output Artifacts

Each scan now creates a dedicated run folder:

• evidence/run_<timestamp>/report.json
• evidence/run_<timestamp>/screenshot.png
• evidence/run_<timestamp>/dom.html

Each report.json includes analysis_method:

• llm-ollama
• llm-openai
• heuristic

Run logs are also appended to:

• logs/runs.jsonl (one JSON entry per scan)

---

🔐 Security Note

Docker sandboxing significantly reduces risk compared to running directly on the host, but no sandbox is perfect. For high-risk investigations, use a dedicated VM and isolated network segment.