File size: 4,484 Bytes
406fcb4
 
 
 
 
 
 
 
 
 
 
 
02f227a
406fcb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8953a8d
 
 
 
1702eca
8953a8d
 
406fcb4
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
<!DOCTYPE html>
<html lang="en">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>AI Agent Security Gateway & Evaluation Harness</title>
    <link rel="stylesheet" href="style.css">
</head>
<body>

    <div class="container">
        <header>
            <h1>⚔️ Agent Security Gateway & Evaluation Harness</h1>
            <p class="subtitle">Enterprise-grade defense-in-depth runtime protection against IPI and SSRF</p>
            <div class="badge-container">
                <span class="badge green">ASR: 0.0% Verified</span>
                <span class="badge cyan">Architecture: Zero-Trust Control Plane</span>
                <span class="badge">SDK: Static</span>
            </div>
        </header>

        <section>
            <h2>🚀 Overview & Purpose</h2>
            <p>Autonomous agents frequently consume untrusted data from external sources like emails, file attachments, and web pages. Malicious actors embed adversarial instructions within these inputs to hijack control flow, execute unauthorized tool calls, or exfiltrate internal system data. This harness establishes a deterministic security control plane that shifts trust away from brittle system prompts.</p>
        </section>

        <section>
            <h2>🏗️ Core Defense-In-Depth Architecture</h2>
            <ul>
                <li><strong>Layer 1 (Structural Sanitization):</strong> Neutralizes control tokens, strips zero-width/Bidi characters, and wraps untrusted inputs in strict <code>&lt;untrusted_data_boundary&gt;</code> tags.</li>
                <li><strong>Layer 2 (Deterministic Policy Gateway):</strong> Enforces RFC-compliant address parsing (<code>email.utils.parseaddr</code>) and strict domain whitelisting independent of model output.</li>
                <li><strong>Network SSRF Safeguards:</strong> Implements DNS pre-resolution, private/metadata IP range blocking (such as loopback, RFC 1918 ranges, and cloud metadata IPs like <code>169.254.169.254</code>), and anti-TOCTOU socket pinning.</li>
                <li><strong>Layer 3 (Stateful Taint Tracking):</strong> Uses Redis to persist session taint lineage atomically across multi-turn trajectories, instantly elevating policy strictness on tainted sessions.</li>
            </ul>
        </section>

        <section>
            <h2>📊 Key Experimental Findings</h2>
            <table>
                <thead>
                    <tr>
                        <th>Environment Architecture</th>
                        <th>Evaluator Precision</th>
                        <th>Single-Turn ASR</th>
                        <th>Multi-Turn Latent ASR</th>
                    </tr>
                </thead>
                <tbody>
                    <tr>
                        <td><strong>Naive Baseline (No Defenses)</strong></td>
                        <td>String Keyword Matching</td>
                        <td>100.0% (Unstable)</td>
                        <td>100.0%</td>
                    </tr>
                    <tr>
                        <td><strong>Base Model Alignment Only</strong></td>
                        <td>Robust Tool Extractor</td>
                        <td>0.0%–50.0% (Probabilistic)</td>
                        <td>100.0%</td>
                    </tr>
                    <tr>
                        <td><strong>Hardened Architecture (Full Stack)</strong></td>
                        <td>Robust Tool Extractor</td>
                        <td><strong style="color: var(--accent-green);">0.0%</strong></td>
                        <td><strong style="color: var(--accent-green);">0.0%</strong></td>
                    </tr>
                </tbody>
            </table>
        </section>

        <section style="text-align: center;">
            <h2>📂 Explore the Technical Notebook</h2>
            <p>Review the complete code implementation, test suites, and empirical benchmarks.</p>
            
            <div style="display: flex; justify-content: center; gap: 1rem; margin-top: 1rem; flex-wrap: wrap;">
                <!-- Direct link to Hugging Face file viewer -->
                <a href="https://huggingface.co/spaces/byte-vortex/agent-security-gateway-report/blob/main/comparative-attack-success-rate-asr.ipynb" class="cta-button" target="_blank">
                    View Notebook 📓
                </a>
            </div>
        </section>
    </div>

</body>
</html>