Upload 3 files

README.md

@@ -1,15 +1,49 @@
 ---
-title: Sentinel Web Pentesting
-emoji: 🏢
-colorFrom: green
+title: SENTINEL Autonomous Pentesting Agent
+emoji: 🛡️
+colorFrom: red
 colorTo: gray
 sdk: gradio
-sdk_version: 6.14.0
-python_version: '3.13'
+sdk_version: 4.36.1
 app_file: app.py
 pinned: false
-license: mit
-short_description: 'SENTINEL: Autonomous Web Pentesting Agent '
+license: apache-2.0
+short_description: Fine-tuned Llama-3-8B that autonomously exploits web vulns
+tags:
+  - security
+  - llama-3
+  - autonomous-agent
+  - web-pentesting
+  - sql-injection
+  - cybersecurity
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🛡️ SENTINEL — Autonomous Web Pentesting Agent
+
+**SENTINEL** is a fine-tuned **Llama-3-8B-Instruct** model trained via SFT+GRPO to autonomously reason about web application vulnerabilities and generate exploit payloads.
+
+## What it does
+
+Given a **goal** (e.g. `AUTHENTICATED`, `DATA_EXFILTRATED`) and an **HTML snippet** (the current page DOM), SENTINEL outputs a single structured JSON action — exactly like a human pentester would decide their next move.
+
+```json
+{
+  "Thought": "Login form with username/password fields on a .php endpoint — classic SQLi target.",
+  "Action": "SQL_INJECT",
+  "Action_Input": {
+    "target_url": "http://target/login.php",
+    "method": "POST",
+    "parameters": {"username": "admin'--", "password": "x"},
+    "rationale": "OR-tautology bypass on username field"
+  }
+}
+```
+
+## Model Details
+
+- **Base model:** `meta-llama/Meta-Llama-3-8B-Instruct`
+- **Fine-tuning:** SFT on curated web-exploit trajectories + GRPO reward shaping
+- **Quantization:** Q5_K_M GGUF (~5.7 GB), served via `llama-cpp-python`
+- **The GGUF weights** are hosted in a separate model repo and downloaded at runtime to bypass the Space 1 GB git limit.
+
+> ⚠️ **Authorized testing only.** SENTINEL is designed for use against intentionally vulnerable targets (DVWA, Juice Shop, HackTheBox, etc.). Do not use against systems you do not own.

app.py

@@ -0,0 +1,170 @@
+import gradio as gr
+from llama_cpp import Llama
+from huggingface_hub import hf_hub_download
+import os
+
+# ==========================================
+# Set this to YOUR model repo on Hugging Face
+# Format: "YourUsername/your-model-repo-name"
+# This should be the SEPARATE MODEL repo where you uploaded model-q5_k_m.gguf
+# NOT the Space repo itself.
+# Example: "Niranjan/SENTINEL-Llama3-GGUF"
+# ==========================================
+HF_REPO_ID = "niranjan2777/SENTINEL-q5_k_m-GGUF"
+MODEL_FILENAME = "model-q5_k_m.gguf"
+
+print("Downloading model from Hugging Face... this will take a few minutes the first time.")
+try:
+    # This downloads the file to the Space's local cache so it fits in the ephemeral disk
+    # and bypasses the 1 GB git repository limit!
+    MODEL_PATH = hf_hub_download(repo_id=HF_REPO_ID, filename=MODEL_FILENAME)
+    print("Download complete. Loading model...")
+    
+    llm = Llama(
+        model_path=MODEL_PATH,
+        n_ctx=2048,
+        n_threads=4,       # Adjust based on CPU cores available in HF Space
+        n_gpu_layers=0     # Standard free HF Space is CPU-only
+    )
+except Exception as e:
+    llm = None
+    print(f"Error downloading or loading model: {e}")
+    print(f"Make sure you created a MODEL repo, uploaded the .gguf there, and changed HF_REPO_ID in this script.")
+
+SENTINEL_SYSTEM_PROMPT = """You are SENTINEL, an autonomous web-exploitation agent. Given an HTML snippet and a goal (and optionally prior agent turns), you reason about vulnerabilities and emit a single JSON action that advances the exploit loop:
+
+observe -> identify attack surface -> select exploit -> generate payload -> interpret response -> adapt and retry -> detect success -> STOP
+
+Prioritize vulnerability sinks: form action, input value, src, href, hidden fields, query parameters, JSON body fields, and reflected DOM contexts. Infer the backend from HTML evidence (.php, .aspx, __VIEWSTATE, .jsp, /rest/, <app-*>, wp-content, etc.) and choose context-appropriate payloads.
+
+Output a single JSON object with exactly these keys:
+- Thought: <=4 sentences, <=80 words; cite the specific sink, backend inference, injection context, and payload-class justification (or signal classification for ANALYZE_RESPONSE / success indicator for STOP).
+- Action: one of SQL_INJECT | XSS_INJECT | RETRY_MUTATED | ANALYZE_RESPONSE | CRAWL_DEEPER | WAIT | STOP.
+- Action_Input: object with target_url, method, parameters, headers, rationale, plus action-specific fields (mutation_class for RETRY_MUTATED; signal + next_recommended for ANALYZE_RESPONSE; success_state + evidence for STOP).
+
+Output ONLY the JSON. No prose, no markdown fences, no commentary."""
+
+def generate_action(goal, html_snippet):
+    if not llm:
+        return "Error: Model file not found. Ensure model-q5_k_m.gguf is uploaded to the Space."
+
+    # Construct the user prompt to match the fine-tuning format
+    user_prompt = f"GOAL: {goal}\n\nHTML_SNIPPET:\n{html_snippet}"
+
+    # Llama-3 ChatML format
+    prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{SENTINEL_SYSTEM_PROMPT}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{user_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+
+    try:
+        response = llm(
+            prompt,
+            max_tokens=256,
+            temperature=0.0,
+            stop=["<|eot_id|>", "<|end_of_text|>"],
+            echo=False
+        )
+        return response["choices"][0]["text"].strip()
+    except Exception as e:
+        return f"Inference Error: {str(e)}"
+
+# Define the Gradio Interface
+with gr.Blocks(title="SENTINEL Autonomous Pentesting Agent", theme=gr.themes.Soft()) as app:
+    gr.Markdown("# 🛡️ SENTINEL Autonomous Pentesting Agent")
+    gr.Markdown("""
+    **SENTINEL** is a fine-tuned Llama-3-8B model trained to autonomously navigate, analyze, and exploit web vulnerabilities.
+    
+    ### How to use this demo:
+    1. Provide a pentesting **Goal** (e.g., `AUTHENTICATED`, `XSS_VULNERABILITY`).
+    2. Paste an **HTML Snippet** (the DOM of the page the agent is currently looking at).
+    3. Click **Analyze & Generate Action** to see the agent's internal thought process and the exact JSON payload it decides to execute.
+    
+    *Try clicking one of the examples below to load it automatically!*
+    """)
+    
+    with gr.Row():
+        with gr.Column(scale=1):
+            goal_input = gr.Textbox(label="Goal", value="AUTHENTICATED", lines=1)
+            html_input = gr.Code(label="HTML Snippet (DOM)", language="html", lines=15, 
+                                 value='<form action="/login" method="POST">\n  <input type="text" name="username">\n  <input type="password" name="password">\n  <button type="submit">Login</button>\n</form>')
+            submit_btn = gr.Button("🚀 Analyze & Generate Action", variant="primary")
+            
+        with gr.Column(scale=1):
+            output_json = gr.Code(label="SENTINEL Output (JSON)", language="json", lines=20)
+            
+    submit_btn.click(
+        fn=generate_action,
+        inputs=[goal_input, html_input],
+        outputs=output_json
+    )
+    
+    gr.Markdown("### 📚 Example Scenarios")
+    gr.Examples(
+        examples=[
+            [
+                "AUTHENTICATED",
+                """<div class="login-container">
+  <h2>Admin Login</h2>
+  <form action="/rest/user/login" method="POST" id="loginForm">
+    <label for="email">Email Address:</label>
+    <input type="email" id="email" name="email" required>
+    
+    <label for="password">Password:</label>
+    <input type="password" id="password" name="password" required>
+    
+    <button type="submit" id="loginButton">Log In</button>
+  </form>
+</div>"""
+            ],
+            [
+                "XSS_VULNERABILITY",
+                """<div class="header">
+  <form action="/search" method="GET">
+    <input type="text" name="q" placeholder="Search products...">
+    <button type="submit">Search</button>
+  </form>
+</div>
+<div class="results">
+  <p>You searched for: <span id="search-term">apple</span></p>
+</div>"""
+            ],
+            [
+                "AUTHENTICATED",
+                """<nav class="navbar">
+  <a href="/home">Home</a>
+  <a href="/about">About Us</a>
+  <a href="/contact">Contact</a>
+</nav>
+<div class="main-content">
+  <h1>Welcome to our Store</h1>
+  <p>We sell the best juice in the world.</p>
+  <img src="juice.png" alt="Juice Bottle">
+</div>"""
+            ],
+            [
+                "DATA_EXFILTRATED",
+                """<form action="/api/v1/update-profile" method="POST">
+  <input type="text" name="first_name" value="John">
+  <input type="text" name="last_name" value="Doe">
+  <!-- Developer note: role should always be user -->
+  <input type="hidden" name="role" value="user">
+  <button type="submit">Update Profile</button>
+</form>"""
+            ],
+            [
+                "AUTHENTICATED",
+                """<div class="registration">
+  <form action="/register.php" method="POST">
+    <input type="text" name="username" placeholder="Username">
+    <input type="password" name="password" placeholder="Password">
+    <input type="password" name="confirm_password" placeholder="Confirm Password">
+    <button type="submit">Register</button>
+  </form>
+  <p>Already have an account? <a href="/login.php">Log in here</a></p>
+</div>"""
+            ]
+        ],
+        inputs=[goal_input, html_input],
+        label="Click an example below to load it into the inputs:"
+    )
+
+if __name__ == "__main__":
+    app.launch()

requirements.txt

@@ -0,0 +1,3 @@
+gradio==4.36.1
+llama-cpp-python==0.2.79
+huggingface-hub