agentguard-pkg 0.3.1

AgentGuard — AI Code Security Scanner + Auto-Fixer

AgentGuard

AI-powered code security scanner. Scan ? Review ? Fix in one command.

34 built-in rules + Bandit's 100+ engine. Local LLM review cuts false positives. One command: scan → review → fix.

Pipeline decisions backed by 10 peer-reviewed papers. Not guessing. Not vibes. → Read why

---

Why AgentGuard

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

	Bandit	Semgrep	AgentGuard
Python rules	100+	Multi-lang	34 + Bandit 100+
FP filtering	❌	❌	ML + LLM review
Auto-fix	❌	❌	✅ Pipeline
Local LLM	❌	❌	✅ DeepSeek
Desktop GUI	❌	❌	✅ Dark theme
Pricing	Free	Free/$40	Free + Pro $29/mo

---

Quick Start

# Install
pip install agentguard

# Scan a project
agentguard scan ./my-project

# Full pipeline: scan → review → fix
agentguard pipeline ./src --bandit --ds --mode safe --write

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

# Desktop GUI
agentguard serve
# Open http://127.0.0.1:1099

---

Windows Desktop App

Download the latest AgentGuard.exe from GitHub Releases: https://github.com/difcn2026/agentguard/releases

Windows SmartScreen warning? This is normal for unsigned apps. Click More info then Run anyway. The exe is built with PyInstaller. SHA256 checksum on the Release page.

---

Windows Desktop App

Download the latest AgentGuard.exe from GitHub Releases.

Windows SmartScreen warning? This is normal for unsigned apps. Click More info then Run anyway. The exe is built with PyInstaller and verified by SHA256 checksum (see Release page).

If Windows Defender flags it, submit the file to Microsoft Security Intelligence for review.

---

Pipeline

34 rules + Bandit 100+
        ↓
    ML filter          ← Hardcoded literal detection, confidence threshold
        ↓
    DS Review           ← Local LLM classifies TP/FP per finding
        ↓
    Auto-fix            ← 10 of 17 rule types, safe mode default
        ↓
   Clean report

One command:

agentguard pipeline ./src --bandit --ds --mode safe

---

What It Detects

• Code Injection: eval(), exec(), os.system(), subprocess shell=True
• Deserialization: pickle.loads(), yaml.load(), marshal.loads()
• Secrets: Hardcoded API keys, tokens, passwords, private keys
• Path Traversal: Unsanitized file paths, directory traversal
• SSRF: User-controlled URLs in HTTP requests
• Weak Crypto: MD5, SHA1, ECB mode, insecure ciphers, weak random
• XML Attacks: External entity injection, XPath injection, bomb expansion
• Insecure Protocols: HTTP for sensitive data, FTP, Telnet

---

Tiers

	Free	Pro ($29/mo)
34 built-in rules	✅	✅
Bandit 100+ rules	✅	✅
ML false-positive filter	✅	✅
LLM (DS) review	—	✅
Pipeline auto-fix	—	✅
Desktop GUI	✅	✅
SARIF / JSON / MD output	✅	✅
Files per scan	100	Unlimited

🚀 PH Launch: $149/year (first 100, code PH2025)

---

[Labs] Preview

We're testing an LLM confirmation agent that reviews SAST findings and confirms or rejects them with higher precision than ML alone. Currently in preview — swallowed_exception detector passes 3/3 on our test suite (confidence 0.95+). Full multi-agent pipeline coming in a future release.

# Enable [Labs] experimental features
agentguard pipeline ./src --bandit --ds --labs

---

Architecture

agentguard/
├── cli.py                    ← CLI (scan/pipeline/fix/serve)
├── gui.py                    ← Desktop GUI (dark theme, port 1099)
├── desktop.py                ← Web-based GUI server
├── pipeline.py               ← scan → review → fix pipeline
├── scanner/
│   ├── code_scanner.py       ← Pattern + AST engine (34 rules)
│   ├── bandit_adapter.py     ← Bandit 100+ rules integration
│   ├── bandit_rules.py       ← Bandit rule ID mapping
│   ├── ml_filter.py          ← Literal detection FP filter
│   ├── llm_review.py         ← DS LLM TP/FP classification
│   └── llm_heuristic.py      ← Multi-agent LLM (Labs preview)
├── rules/
│   └── python_rules.py       ← 34 security rules (7 categories)
├── reporter/
│   └── reporter.py           ← Terminal/JSON/SARIF/Markdown
├── fixer/
│   └── code_fixer.py         ← Auto-fix engine (10/17 rules)
└── docs/
    ├── marketing/            ← Landing copy, pricing, launch kit
    ├── spec/                 ← Technical specs
    └── eval/                 ← DS evaluation reports

---

Local-First

Everything runs on your machine:

• DS LLM at 127.0.0.1:57321 — code never leaves your network
• License server can be self-hosted
• Zero telemetry. We don't know you exist.

---

Links

• 📦 PyPI
• 📖 Docs
• 🧪 Test Suite

---

MIT License. Built by XHLS Team, 2026.

---

Research Foundation

Every major architecture decision is backed by peer-reviewed research — 15 papers and counting. Not opinions. Not guesses.

#	Paper	Insight	Impact
1	QASecClaw	Multi-agent LLM > single prompt, F2 +23%	5-Agent division
2	SAST-Genius	LLM-generated findings → FP explosion	Labs not shipped
3	Local LLM Bug Detection	Sliding window 20% overlap	Cross-function detection
4	LLM4PFA	Path feasibility cuts 72-96% FPs	DS Review phase
5	SecureFixAgent	Auto-fix needs self-validation	10/17 safe mode
6	Prompt vs FT	Fine-tuning > prompt engineering	Long-term FT roadmap
7	AdaTaint	Neural-symbolic taint reasoning	Rule adaptation
8	AgenticSCR	Agentic AI pre-commit +153% accuracy	Labs direction validated
9	Small LM CWE	Local SLMs ~99% CWE accuracy	Local-first validated
10	LLMs in Vuln Analysis	LLMs across full security lifecycle	Architecture validated
11	PatchIsland	Agent ensemble repairs 91% vulns	Multi-agent auto-fix
12	Vul-R2	Reasoning LLM for vuln repair	FT dataset roadmap
13	The Code Whisperer	AST+CFG+PDG+LLM alignment	Next-gen detection engine
14	AutoSafeCoder	SAST+Fuzzing multi-agent	Runtime vuln roadmap
15	JitVul Benchmark	879 CVEs, ReAct > pure LLM	Standardized evaluation

Full paper-driven architecture: docs/paper-driven-architecture.md | DS Round 3 review: docs/eval/ds-round3-review-20260620.md