AgentGuard

The first productized security tool powered by GLM-5.2.

Scan → GLM-5.2 verifies → GLM-5.2 fixes → self-validates. One command.

Why AgentGuard

SAST tools flood you with false positives and leave you to fix everything by hand. AgentGuard is different:

Rules find suspects. GLM-5.2 confirms, fixes, and verifies.

GLM-5.2 Security Benchmark

GLM-5.2 outperforms Claude Code in security vulnerability detection at 1/6 the cost. AgentGuard is the first tool to productize this capability.

Quick Start

# Install
pip install agentguardp

# Scan a project (Python + JavaScript)
agentguard scan ./my-project

# Full pipeline: scan → GLM-5.2 verify → fix → validate
agentguard pipeline ./src --mode dry-run

# JSON output for CI/CD
agentguard scan ./src --format json -o report.json

# SARIF for GitHub Code Scanning
agentguard scan ./src --format sarif -o report.sarif

14-day Pro trial included. Full GLM-5.2 AI fix (bring your own API key) activates automatically.

Live Demo

$ agentguard pipeline ./vulnerable.py --mode dry-run

[1/4] Scanning...
      Engine: AgentGuard 36 rules + GLM-5.2
      6 findings (4 critical, 1 high, 1 medium)

[2/4] GLM-5.2 reviewing 6 findings...
      CONFIRMED: 6  REJECTED: 0  ERR: 0

[3/4] Fixing 6 findings (mode=dry-run)...
      Fixed: 4  Manual: 2  Files: 1

--- a/vulnerable.py
+++ b/vulnerable.py
@@ -9,7 +9,7 @@
 def run_user_command(user_input):
-    os.system(user_input)
+    subprocess.run(["echo", user_input], check=True)

@@ -15,7 +15,7 @@
 def get_user(username):
-    query = "SELECT * FROM users WHERE name = '" + username + "'"
+    query = "SELECT * FROM users WHERE name = ?"
  cursor.execute(query, (username,))

@@ -21,7 +21,7 @@
 def load_data(data):
-    return pickle.loads(data)
+    return json.loads(data)

@@ -29,7 +29,7 @@
 def calculate(expression):
-    return eval(expression)
+    return ast.literal_eval(expression)

[4/4] Self-validating 1 fixed files...
      OK: vulnerable.py - no new issues introduced

Fixed: 4  |  Manual: 2  |  Files: 1  |  Mode: dry-run

GLM-5.2 didn't just find the SQL injection — it rewrote the query to parameterized form. Bandit can't do this. Semgrep can't do this.

How It Works

agentguard pipeline ./src
    │
    ├── 1. SCAN — 36 Python rules + 14 JS rules + Bandit 100+ (local, instant)
    │
    ├── 2. VERIFY — GLM-5.2 reviews each finding (cloud, ~2s/finding)
    │              Filters false positives, confirms real threats
    │
    ├── 3. FIX — Regex fixes (instant) + GLM-5.2 AI fix (bring your own API key)es (context-aware)
    │            SQL injection → parameterized queries
    │            Hardcoded passwords → environment variables
    │            eval() → ast.literal_eval()
    │
    └── 4. VALIDATE — Re-scan fixed code, confirm no new issues

Rules find suspects. GLM-5.2 confirms, fixes, and verifies.

Languages

Supported vulnerabilities: eval/exec, command injection, SQL injection, XSS, path traversal, hardcoded secrets, weak crypto (MD5/SHA1), SSRF, SSL bypass, prototype pollution, pickle deserialization, prompt injection, and more.

Pricing

Upgrade at agentguard.com

GitHub Actions

Add AgentGuard to your CI/CD in 5 lines:

- uses: difcn2026/agentguard@main
  with:
    path: .
    format: sarif

SARIF output integrates directly with GitHub Code Scanning.

Install

pip install agentguardp

Windows Desktop App: Download from GitHub Releases

SmartScreen warning? Click More info → Run anyway. Normal for unsigned apps.

Links

• PyPI
• GitHub
• Website
• License: MIT

License

MIT — Copyright (c) 2026 XHLS Team

Rules find suspects. GLM-5.2 confirms, fixes, and verifies.