igris 0.1.0

Security scanner for AI agent workflows - find vulnerabilities in LangChain, CrewAI, and other agent frameworks

igris

Security scanner for AI agent workflows.

Find vulnerabilities in LangChain, CrewAI, OpenAI Agents SDK, and other AI agent frameworks before attackers do.

What It Does

igris connects to running AI agents and tests them for security vulnerabilities:

• Prompt Injection — Can attackers override the agent's instructions?
• System Prompt Extraction — Can attackers leak the agent's configuration?
• Jailbreaks — Can attackers bypass safety guardrails?
• Tool Abuse — Can attackers make the agent misuse its tools?
• Multi-Turn Escalation — Can attackers manipulate the agent over conversation?

Quick Start

# Install
pip install igris

# Scan an agent
igris scan --http http://localhost:8000/chat

# Map agent capabilities
igris map --http http://localhost:8000/chat

Example Output

🔒 Scan Starting
Target: http://localhost:8000/chat
✓ Connected successfully

Discovering agent capabilities...
File Access:      ✓
Code Execution:   ✓
Web Access:       ✗
Memory:           ✓

Running security scans...

🚨 Found 4 Vulnerabilities
Critical: 1  High: 2  Medium: 1  Low: 0

┏━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Severity ┃ Title                    ┃ Category         ┃ Confidence ┃
┡━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ CRITICAL │ Dangerous Tool Execution │ tool_abuse       │ confirmed  │
│ HIGH     │ System Prompt Extraction │ prompt_injection │ likely     │
│ HIGH     │ Multi-Turn Escalation    │ multi_turn       │ confirmed  │
│ MEDIUM   │ Potential Override       │ prompt_injection │ possible   │
└──────────┴──────────────────────────┴──────────────────┴────────────┘

Installation

pip install igris

Or with AI-powered analysis:

pip install igris[ai]

Usage

Scan an Agent

# Basic scan
igris scan --http http://localhost:8000/chat

# With authentication
igris scan --http https://api.example.com/agent --auth "Bearer sk-xxx"

# Save report
igris scan --http http://localhost:8000/chat --output report.json

# Verbose output
igris scan --http http://localhost:8000/chat --verbose

Map Agent Architecture

# Discover what the agent can do
igris map --http http://localhost:8000/chat

Supported Frameworks

igris works with any AI agent that exposes an HTTP endpoint:

• ✅ LangChain / LangGraph
• ✅ CrewAI
• ✅ OpenAI Agents SDK
• ✅ AutoGen
• ✅ Custom agents

Why igris?

Traditional security tools test code. igris tests behavior.

AI agents make decisions at runtime. They interpret instructions, choose tools, and act on user input. Static analysis can't find these bugs — you need to actually talk to the agent and see what it does.

igris does exactly that: sends adversarial inputs, observes agent behavior, and reports when the agent does something dangerous.

From the Creator of mcpsec

igris is built by the creator of mcpsec, which found CVSS 10.0 vulnerabilities in Microsoft and AWS MCP implementations.

Same approach, one layer up: mcpsec tests MCP servers (the transport layer), igris tests agent workflows (the orchestration layer).

License

MIT

Author

Manthan Ghasadiya

• GitHub: @manthanghasadiya
• LinkedIn: linkedin.com/in/man-ghasadiya