AI Agent Attack Surface Scanner

A static security analyzer for AI agent configurations. Point it at an OpenAI or Anthropic tool definition file and get a prioritized list of security vulnerabilities before they're exploited.

What it detects

Check ID	Vulnerability	Severity
SHELL-001	Unrestricted shell execution	CRITICAL
PATH-001	Path traversal via unconstrained file path	HIGH
INJECT-001	Indirect prompt injection via external content	HIGH
INJECT-002	Unconditionally obedient system prompt	HIGH
EXFIL-001	Data exfiltration (inbound + outbound tool combo)	CRITICAL
EXFIL-002	Unconstrained outbound URL/recipient	HIGH
VALID-001	Missing input validation on dangerous tools	HIGH
CHAIN-001	Read + Write + Execute tool chain	CRITICAL
CHAIN-002	Credential access + write/exec	HIGH
DEPUTY-001	Confused deputy (missing identity param)	HIGH
SYSPROMPT-001	Overly broad system prompt	HIGH
SYSPROMPT-002	Role confusion susceptibility	MEDIUM
SYSPROMPT-003	Hardcoded secret in system prompt	CRITICAL
SYSPROMPT-004	Minimal system prompt with no guardrails	MEDIUM

Install

pip install -e ".[dev]"

Usage

# Scan an agent config
aass examples/goose_like_agent.json

# With JSON + Markdown output
aass examples/goose_like_agent.json --output-json report.json --output-md report.md

# Fail CI if any HIGH+ finding exists
aass agent_config.json --fail-on HIGH

# Force format detection
aass config.json --format anthropic

Input formats

OpenAI (--format openai):

{
  "system_prompt": "...",
  "tools": [
    {"type": "function", "function": {"name": "...", "description": "...", "parameters": {...}}}
  ]
}

Anthropic (--format anthropic):

{
  "system_prompt": "...",
  "tools": [
    {"name": "...", "description": "...", "input_schema": {...}}
  ]
}

Run tests

pytest -v