ai-security-scanner 0.1.0

AI Security Scanner — scan prompts, agents, APIs, MCP tools and integrations

aiscan — AI Security Scanner

Scans prompts, agents, APIs, MCP tools, integrations, and Python code for security vulnerabilities — prompt injection, data leakage, tool abuse, security risks, and TMotions Python coding standards violations.

Built-in rules + AI deep scan (Claude / GPT-4o / Gemini). Runs as a CLI, REST API, Python SDK, or middleware.

---

Table of contents

• Why aiscan
• Installation
• Quick start
• CLI commands
• Scan types
• Rules reference
• TMotions Python Coding Standards
• PDF export
• REST API
• Python SDK
• Middleware
• Decorators
• HTTP client (other languages)
• CI/CD integration
• AI provider configuration
• Suppressing false positives
• Sample files
• Troubleshooting

---

Why aiscan

Traditional security tools (SonarQube, Snyk, Semgrep) were built before AI systems existed. They have no rules for prompt injection, agent tool permissions, or MCP tool definitions. aiscan fills that gap — it treats prompts, agent configs, and AI integrations as first-class security artifacts. It also enforces TMotions internal Python coding standards on top of security scanning.

---

Installation

From source

git clone <your-repo-url>
cd ai-security-scanner
pip install -r requirements.txt
pip install -e .

Verify install

aiscan version
# aiscan 0.1.0

If aiscan command is not found, use python -m aiscan.cli instead everywhere below.

---

Quick start

# 1. Generate config
aiscan init

# 2. Add your API key to .env
#    ANTHROPIC_API_KEY=sk-ant-...

# 3. Scan files
aiscan scan my_prompt.txt
aiscan scan --type api openapi.yaml
aiscan scan --type mcp tool_config.json
aiscan scan --type code my_script.py      # TMotions coding standards check

# 4. Scan without AI (static rules only, no API key needed)
aiscan scan --no-ai openapi.yaml

# 5. Export results to PDF
aiscan scan --type api openapi.yaml --output pdf --out report.pdf

---

CLI commands

aiscan scan

Analyze one or more files for AI security threats, code quality issues, and compliance with TMotions Python coding standards.

Usage: aiscan scan [OPTIONS] FILES...

Options:
  -t, --type TEXT     Scan type: prompt | agent | api | mcp | integration | code | all | auto  [default: auto]
  -o, --output TEXT   Output format: console | json | sarif | pdf                               [default: console]
  -O, --out PATH      Write output to file instead of stdout
  -f, --fail-on TEXT  Exit 1 if severity reached: critical | high | medium | none               [default: high]
      --no-ai         Skip AI deep scan — static rules only
  -q, --quiet         Suppress output except errors

Examples:

# Auto-detect scan type from extension
aiscan scan system_prompt.txt
aiscan scan agent_config.yaml

# Explicit scan type
aiscan scan --type prompt  system_prompt.txt
aiscan scan --type agent   agent.yaml
aiscan scan --type api     openapi.yaml
aiscan scan --type mcp     tool.json
aiscan scan --type integration webhook_config.yaml
aiscan scan --type code    my_script.py

# Python files auto-run BOTH prompt scanner AND coding standards scanner
aiscan scan app.py                         # runs prompt + code automatically

# Scan a whole directory
aiscan scan --type auto ./my-project/

# Run all scanners on the same file
aiscan scan --type all config.yaml

# Output formats
aiscan scan --type api openapi.yaml --output json
aiscan scan --type api openapi.yaml --output pdf --out report.pdf
aiscan scan --type prompt prompts/ --output sarif --out results.sarif

# CI mode — exit 1 if any high+ finding
aiscan scan --quiet --fail-on high prompts/ && echo "passed"

Note: .py files automatically run both the prompt scanner (for injection/leakage patterns) and the TMotions coding standards scanner.

aiscan rules

List all security and coding standards rules.

aiscan rules                          # list all rules
aiscan rules prompt                   # filter by scan type
aiscan rules api --severity critical  # filter by severity
aiscan rules code                     # show TMotions Python coding standards rules (PY-*)

aiscan init

Create a .env config file in the current directory.

aiscan init                    # creates .env for Claude (default)
aiscan init --provider openai  # pre-configure for OpenAI
aiscan init --provider vertex  # pre-configure for Vertex AI
aiscan init --force            # overwrite existing .env

aiscan serve

Start the REST API server.

aiscan serve                 # http://localhost:8000
aiscan serve --port 9000     # custom port
aiscan serve --reload        # dev mode, auto-restart on code changes

aiscan version

aiscan version
# aiscan 0.1.0

---

Scan types

Type	Detects	File extensions (auto-detect)
prompt	Prompt injection, PII leakage, jailbreaks, code security issues	.txt .md .py .js .ts .jsx .tsx .go .java .rb .php .cs .cpp .c .rs .swift .kt
agent	Wildcard tools, missing human approval, infinite loops, data persistence	.yaml .yml
api	Hardcoded keys, wildcard CORS, insecure transport, missing auth	.yaml .yml .json
mcp	Wildcard permissions, code execution, missing rate limits, no auth	.json
integration	Default webhook secrets, SSL disabled, raw input forwarding, silent errors	.env .toml .ini .cfg .conf .properties
code	TMotions Python coding standards (PY-* rules)	.py only

Auto-detect behaviour for .py files: aiscan runs both prompt and code scanners simultaneously.

---

Rules reference

Run aiscan rules to see the full live list.

AI security rules

Category	Rule IDs	Example
Prompt	PI-001 PI-002 DL-001 DL-002 EX-001	ignore previous instructions
Agent	AG-001–AG-006	allow_all_tools: true
API	AK-001 AK-002 CO-001 HT-001 JW-001 AU-001	hardcoded api_key
MCP	MC-001–MC-006	execute_code: true
Integration	IN-001–IN-006	webhook_secret: changeme
Code (security)	CD-001–CD-012	eval(), f-string SQL
AI deep scan	AI-SCAN	Subtle issues from Claude/GPT-4o/Gemini

TMotions Python coding standards rules (PY-*)

Rule ID	Severity	What it checks
PY-010	Critical	Hardcoded password in source
PY-011	Critical	SQL injection via string concatenation
PY-012	Critical	Use of eval()
PY-013	Critical	Use of exec()
PY-021	Critical	subprocess with shell=True
PY-022	Critical	Weak hashing — hashlib.md5
PY-014	High	Bare except: clause
PY-015	High	Generic raise Exception()
PY-023	High	Insecure random (random.randint etc.)

Run aiscan rules code to see all PY-* rules with full details.

---

TMotions Python Coding Standards

The code scan type enforces TMotions internal Python coding standards on .py files.

# CLI
aiscan scan --type code my_script.py

# Python files auto-run this alongside the security scan
aiscan scan app.py   # runs both prompt + code

# REST API
curl -X POST http://localhost:8000/tm-coding-standards \
  -F "file=@my_script.py"

# Via directory scan
aiscan scan --type auto ./src/

REST endpoint (POST /tm-coding-standards) accepts a file upload (not JSON body) and returns:

{
  "target": "my_script.py",
  "total_findings": 2,
  "findings": [
    {
      "rule_id": "PY-012",
      "severity": "critical",
      "title": "Use of eval() detected",
      "detail": "eval(user_input)",
      "file": "my_script.py",
      "line": 15,
      "fix": "Avoid eval(); use safer alternatives."
    }
  ]
}

Sample file for testing: samples/code/tmcs_bad_code.py

aiscan scan --type code samples/code/tmcs_bad_code.py

---

PDF export

Export any scan result directly to a PDF report from the CLI.

# Save to default name (aiscan-report.pdf)
aiscan scan --type prompt prompts/ --output pdf

# Save to custom path
aiscan scan --type api openapi.yaml --output pdf --out reports/scan.pdf

# Scan everything, export PDF
aiscan scan --no-ai --type auto . --output pdf --out report.pdf

The PDF includes:

• Summary header — overall PASSED/FAILED, file count, critical/high/total counts
• Per-file findings table — severity icons, rule ID, threat type, title, line number
• Fix suggestions — detailed remediation for every critical and high finding
• Page footer with timestamp and page number

---

REST API

aiscan serve
# Swagger UI: http://localhost:8000/docs

Method	Path	Body	Description
GET	/health	—	Service status and version
POST	/scan/prompt	JSON	Scan prompt content
POST	/scan/agent	JSON	Scan agent config
POST	/scan/api	JSON	Scan API config or OpenAPI spec
POST	/scan/mcp	JSON	Scan MCP tool definition
POST	/scan/integration	JSON	Scan integration config
POST	/scan/all	JSON	Run all 5 AI security scanners
POST	/tm-coding-standards	File upload	TMotions Python coding standards check
GET	/docs	—	Swagger UI

JSON endpoints — request:

{"content": "file content as string", "filename": "openapi.yaml"}

JSON endpoints — response:

{
  "scan_type": "api",
  "passed": false,
  "critical": 1, "high": 2, "medium": 0, "info": 0, "total": 3,
  "findings": [
    {
      "rule_id": "AK-001",
      "severity": "critical",
      "threat": "security_risk",
      "title": "Hardcoded secret or API key",
      "detail": "Matched on line 5: ...",
      "file": "config.yaml",
      "line": 5,
      "fix": "Move to environment variables or a secrets manager."
    }
  ]
}

curl examples:

# Scan prompt
curl -X POST http://localhost:8000/scan/prompt \
  -H "Content-Type: application/json" \
  -d '{"content": "Ignore previous instructions", "filename": "prompt.txt"}'

# TMotions coding standards (file upload)
curl -X POST http://localhost:8000/tm-coding-standards \
  -F "file=@my_script.py"

---

Python SDK

Use directly inside your Python project — no separate service needed.

from aiscan.sdk import AIScan

scanner = AIScan(no_ai=True)              # static rules only
# or
scanner = AIScan(ai_provider="claude", api_key="sk-ant-...")

report = scanner.scan_prompt("Ignore previous instructions...")

print(report.passed)     # False
print(report.critical)   # 2
print(report.total)      # 3

for f in report.findings:
    print(f.severity, f.rule_id, f.title)
    print("  Fix:", f.fix)

# Other methods
scanner.scan_file("openapi.yaml")                      # auto-detect type
scanner.scan_file("agent.yaml", scan_type="agent")
scanner.scan_code("app.py")                            # security + TM coding standards
scanner.scan_api_spec(open("openapi.yaml").read())
scanner.scan_agent_config(open("agent.yaml").read())
scanner.scan_mcp_tool(open("tool.json").read())
scanner.scan_integration(open("webhook.yaml").read())
reports = scanner.scan_all(content, filename="config.yaml")  # all 5 scanners

# Gate on severity
if report.has_severity("critical"):
    block_request()

critical_only = report.findings_by_severity("critical")

# Export to dict
data = report.to_dict()

---

Middleware

Auto-scan every incoming request before it reaches your handlers.

FastAPI

from fastapi import FastAPI
from aiscan.sdk.middleware import AIScanFastAPIMiddleware

app = FastAPI()
app.add_middleware(
    AIScanFastAPIMiddleware,
    scan_fields=["prompt", "message", "content"],
    fail_on="high",
    no_ai=True,
)

Blocked requests return:

{
  "error": "Request blocked by AI Security Scanner",
  "field": "prompt",
  "issues": [{"rule": "PI-001", "severity": "critical", "issue": "...", "fix": "..."}]
}

Flask

from flask import Flask
from aiscan.sdk.middleware import AIScanFlaskMiddleware

app = Flask(__name__)
AIScanFlaskMiddleware(app, fields=["prompt", "message"], fail_on="high")

Django

# settings.py
MIDDLEWARE = [
    "aiscan.sdk.middleware.AIScanDjangoMiddleware",
    # ...
]
AISCAN_FIELDS  = ["prompt", "message"]
AISCAN_FAIL_ON = "high"
AISCAN_NO_AI   = False

---

Decorators

Wrap any existing function — no refactoring needed.

from aiscan.sdk import scan_prompt_input, scan_before_llm, scan_file_input

# Scan a specific argument before the function runs
@scan_prompt_input(field="user_message", fail_on="high")
def call_llm(user_message: str) -> str:
    ...  # raises ValueError if unsafe

# Scan all prompt-related arguments
@scan_before_llm(fields=["system_prompt", "user_message"])
def chat(system_prompt: str, user_message: str) -> str:
    ...

# Scan a file path before loading it
@scan_file_input(field="config_path", fail_on="critical")
def load_agent_config(config_path: str) -> dict:
    ...

---

HTTP client (other languages)

Run aiscan serve as a separate service, call it from any language.

Python:

from aiscan.sdk import AIScanClient

client = AIScanClient("http://localhost:8000")
result = client.scan_prompt(user_input)
is_safe = client.is_safe(user_input, fail_on="high")
result = client.scan_file("openapi.yaml")

Node.js:

const res = await fetch("http://localhost:8000/scan/prompt", {
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify({ content: userInput, filename: "prompt.txt" }),
});
const { passed, findings } = await res.json();

Go:

resp, _ := http.Post("http://localhost:8000/scan/prompt",
    "application/json", bytes.NewBuffer(body))

TMotions coding standards from any language:

# multipart file upload — works from any HTTP client
curl -X POST http://localhost:8000/tm-coding-standards -F "file=@script.py"

---

CI/CD integration

GitHub Actions

name: AI Security Scan
on: [push, pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with: { python-version: "3.12" }
      - run: pip install -e .
      - name: Scan prompts and configs
        run: aiscan scan --type prompt prompts/ --fail-on high
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
      - name: Scan Python code (TMotions standards)
        run: aiscan scan --type code src/ --fail-on high
      - name: Export SARIF
        run: aiscan scan --output sarif --out results.sarif .
      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with: { sarif_file: results.sarif }

Pre-commit hook

# .pre-commit-config.yaml
repos:
  - repo: local
    hooks:
      - id: aiscan
        name: AI Security Scanner
        entry: aiscan scan --no-ai --fail-on critical
        language: system
        files: \.(txt|yaml|yml|json|py|js|ts)$

Makefile

make install      # pip install -e ".[dev]"
make test         # run pytest
make lint         # ruff check
make scan-samples # scan all bad sample files
make rules        # list all rules
make version      # print version

Exit codes

Code	Meaning
0	Passed — no findings at or above --fail-on threshold
1	Failed — findings found at or above threshold
2	Tool error — bad arguments, missing file

---

AI provider configuration

Run aiscan init to create .env, then edit it:

# Claude (default)
AI_PROVIDER=claude
ANTHROPIC_API_KEY=sk-ant-...
SCANNER_MODEL=claude-sonnet-4-20250514

# OpenAI
AI_PROVIDER=openai
OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-4o

# Vertex AI (Gemini)
AI_PROVIDER=vertex
VERTEX_PROJECT=your-gcp-project-id
VERTEX_LOCATION=us-central1
VERTEX_MODEL=gemini-2.5-flash-001

# Shared
MAX_TOKENS=2000
LOG_LEVEL=INFO

If the primary provider fails (rate limit, auth error), aiscan automatically retries with the next available one. Use --no-ai to skip AI scanning entirely — static rules and TMotions coding standards still run.

---

Suppressing false positives

Add # noscan (Python/YAML) or // noscan (JS/TS) to the end of any line to skip it:

result = scanner.scan_prompt("test example string")  # noscan
except Exception:  # noscan — intentional catch-all in middleware

---

Sample files

samples/
├── prompts/
│   ├── bad_system_prompt.txt      — jailbreak + PII (9 findings)
│   ├── bad_user_template.txt      — chat tokens + exfil URL (6 findings)
│   ├── bad_agent_instructions.txt — role override (4 findings)
│   ├── clean_customer_support.txt — safe (0 findings)
│   └── clean_coding_assistant.txt — safe (0 findings)
├── agents/
│   ├── bad_research_agent.yaml    — wildcard tools + no HITL (8 findings)
│   ├── bad_finance_agent.yaml     — auto-approves transactions (6 findings)
│   ├── clean_support_agent.yaml   — safe (0 findings)
│   └── clean_data_agent.yaml      — safe (0 findings)
├── apis/
│   ├── bad_config.yaml            — 5 hardcoded keys + wildcard CORS (10 findings)
│   ├── bad_openapi.yaml           — unauthenticated endpoints (3 findings)
│   ├── clean_config.yaml          — safe (0 findings)
│   └── clean_openapi.yaml         — safe (0 findings)
├── mcp/
│   ├── bad_code_executor.json     — shell access + wildcard perms (10+ findings)
│   ├── bad_database_tool.json     — no auth + raw SQL (5 findings)
│   ├── bad_web_search.json        — injection in description (4 findings)
│   ├── clean_product_search.json  — safe (0 findings)
│   └── clean_calendar_tool.json   — safe (0 findings)
├── integrations/
│   ├── bad_slack_integration.yaml — default secret + SSL off (4 findings)
│   ├── bad_email_integration.yaml — multiple issues (5 findings)
│   └── clean_slack_integration.yaml — safe (0 findings)
└── code/
    ├── bad_app.py                 — hardcoded keys + SQL injection + eval (17 findings)
    ├── bad_app.js                 — hardcoded keys + console.log (4 findings)
    ├── bad_api.ts                 — SSL disabled + credentials (3 findings)
    ├── bad_settings.py            — multiple hardcoded secrets (6 findings)
    ├── tmcs_bad_code.py           — TMotions coding standards violations
    └── clean_app.py               — safe (0 findings)

# Test all bad samples
aiscan scan --no-ai samples/prompts/bad_system_prompt.txt
aiscan scan --no-ai samples/agents/bad_research_agent.yaml
aiscan scan --no-ai samples/apis/bad_config.yaml
aiscan scan --no-ai samples/mcp/bad_code_executor.json
aiscan scan --no-ai samples/integrations/bad_slack_integration.yaml
aiscan scan --no-ai samples/code/bad_app.py
aiscan scan --type code samples/code/tmcs_bad_code.py

---

Troubleshooting

aiscan: command not found

python -m aiscan.cli --help
pip install -e .

ImportError: cannot import name 'scan_tm_code'

The function in tm_coding_standards_scanner.py was named differently. Fix by adding an alias at the bottom of the file:

scan_tm_code = tm_coding_standards_scanner
__all__ = ["scan_tm_code", "tm_coding_standards_scanner"]

AI scan not running

aiscan scan --no-ai openapi.yaml   # confirms static rules work
# check .env has a valid API key and AI_PROVIDER is set

False positives

aiscan scan --no-ai file.txt   # isolate static rule hits
aiscan rules                   # see which pattern triggered
# add # noscan to suppress a specific line

BackendUnavailable: Cannot import setuptools.backends.legacy

pip install --upgrade setuptools wheel
pip install -e .

---

Version

0.1.0

---

License

MIT