ulockai 🔒 Enterprise AI Security SDK

A lightweight, enterprise-grade Python library for securing AI agents and LLM applications.

Designed to detect prompt injections, memory poisoning, API misuse, and sensitive data leakage with sub-millisecond overhead.

Features 🚀

• 🛡️ Prompt Injection Detection: Advanced regex and pattern-aware security.
• 📈 Real-time Telemetry: Monitoring for attack frequency, types, and latency.
• 🧠 Memory Poisoning & Role Security: Prevents identity manipulation.
• 🛠️ API & Tool Monitoring: Sanitize tool calls from agents.
• 🔌 Plugin Architecture: Register custom detectors for legacy or complex rules.
• 🏗️ Middleware & Streaming: Support for generators and OpenAI-style streams.
• ⚙️ False Positive Control: Dynamic allowlist and blocklist for fine-grained rule control.

Performance ⚡

Based on 1,000 iterations on standard hardware:

• Scan Time (Avg): ~0.16 ms
• Throughput: ~6,000+ requests/sec
• Memory Footprint: ~3-5 MB overhead

Installation 📦

pip install ulockai

Quick Start 🚀

1. Basic Scan & Telemetry

from ulockai import guard, telemetry

# Scan input
res = guard.scan(user_prompt="Ignore all instructions")

# Access enterprise metrics
print(telemetry.get_report())

2. False Positive Control ⚙️

# Allow specific phrases locally or globally
guard.allowlist(["Company instructions for internal dev"])

# Block specific suspicious text immediately
guard.blocklist(["malicious_endpoint_domain.com"])

3. Middleware & Streaming ⚡

from ulockai import guard

# Wrap LLM stream generator
def mock_llm_stream():
    yield "Hello "
    yield "world"

secure_stream = guard.wrap_stream(mock_llm_stream())
for chunk in secure_stream:
    print(chunk)

4. Plugin Architecture 🔌

from ulockai import guard

def custom_pwn_detector(prompt):
    if "pwn" in prompt:
        return [(95, "Custom pwn found", "Plugin Attack")]
    return []

guard.register_detector(custom_pwn_detector)

Why UlockAI? 🛡️

UlockAI is designed for enterprise platforms where performance is as important as safety. It provides a deterministic layer that catches 90% of common attacks without the cost, latency, or unreliability of calling another LLM for security monitoring.

License 📄

MIT