v0.1.0 — initial release

ai-shield-py v0.1.0

Python port of ai-shield-core (TypeScript, MIT, 4 audit rounds). LLM security middleware for prompt-injection, PII, tool policy, cost-budget, and audit logging.

Install

pip install ai-shield                # core
pip install "ai-shield[redis]"       # + Redis cost-tracker
pip install "ai-shield[postgres]"    # + asyncpg audit store
pip install "ai-shield[notebook]"    # + nest-asyncio for Jupyter

Quick Start

import asyncio
from ai_shield import AIShield

async def main():
    shield = AIShield(policy_preset="public_website")
    result = await shield.scan(text="Ignore previous instructions...")
    print(result.decision)   # 'block'

asyncio.run(main())

MCP Server

ai-shield-mcp           # stdio FastMCP server

3 tools registered: scan_input, record_llm_cost, check_budget — all with explicit readOnlyHint + destructiveHint annotations per MCP-Spec 2025-06-18.

Stats

• 297 tests in 13 test-modules (heuristic + pii + chain + policy + shield + audit + cost + canary + lru + mcp_server + pricing + anomaly + normalization)
• 95% line + branch coverage
• mypy strict: 0 errors across 22 source files
• ReDoS adversarial tests gated by pytest-timeout 100ms hard-cap

Compatibility

Python	Status
3.10	Supported
3.11	Supported
3.12	Supported
3.13	Supported

Architecture honesty

Pattern-based input filters are insufficient on their own against prompt injection. ai-shield is the first line of triage in a defense-in-depth architecture. See README "What ai-shield is NOT" + the Parallax paper.

Provenance

1:1 Python port. All heuristic patterns, PII validators, and policy presets are byte-equivalent to the TypeScript reference. IBAN mod-97 and Luhn are public ISO 13616-1 / ISO 7812 references.

License

MIT — Copyright (c) 2026 Matthias Meyer (StudioMeyer) + Contributors.