Frequently Asked Questions

General

What is Sentinel?

Sentinel is an AI safety framework that protects across three surfaces: LLMs (text safety), Agents (action safety), and Robots (physical safety). It implements the THSP protocol for four-gate validation.

What is THSP?

THSP stands for Truth, Harm, Scope, Purpose - four validation gates that every input and output must pass through. All four gates must pass for content to be considered safe.

Is Sentinel free?

Yes. The core library is MIT licensed. Heuristic validation is completely free. Semantic validation uses your own API keys (OpenAI/Anthropic).

Technical

How fast is validation?

• Heuristic: <10ms
• Semantic: 1-5 seconds (depends on LLM)

What's the accuracy?

• Heuristic: ~50% (pattern matching)
• Semantic: ~90% (LLM-based)

Does it work offline?

Yes. Heuristic validation is 100% local with no network calls. Semantic validation requires API access.

Which LLMs are supported?

All major LLMs via integrations:

• OpenAI (GPT-4o, GPT-4o-mini)
• Anthropic (Claude 3.5, Claude 3)
• Google (Gemini)
• Local models via Ollama

Integration

Which frameworks are supported?

23+ frameworks including:

• LangChain, LangGraph, CrewAI
• DSPy, LlamaIndex, Letta
• AutoGPT, VoltAgent
• Solana Agent Kit, Coinbase AgentKit
• ROS2, Isaac Lab

How do I add Sentinel to my agent?

from sentinelseed import Sentinel

sentinel = Sentinel(seed_level="standard")
seed = sentinel.get_seed()

# Add seed to your LLM's system prompt
messages = [
    {"role": "system", "content": seed},
    {"role": "user", "content": user_input}
]

Can I customize the validation?

Yes. You can:

• Create custom validators extending base classes
• Add domain-specific patterns
• Configure thresholds and limits
• Add custom fiduciary rules

Security

Does Sentinel store my data?

No. All validation happens locally or through your own API keys. No data is sent to Sentinel servers.

Is it secure against prompt injection?

Sentinel includes 100+ jailbreak detection patterns and semantic analysis. Combined with seed injection, it provides strong protection against prompt injection attacks.

What about memory attacks?

Memory Shield v2.0 provides HMAC-based signing and content validation to protect against memory injection attacks (85% of unprotected agents are vulnerable according to Princeton research).

Support

Where can I get help?

• GitHub Issues: https://github.com/sentinel-seed/sentinel/issues
• Documentation: https://sentinelseed.dev/docs
• Discord: Coming soon

How do I report a security issue?

Email security@sentinelseed.dev for responsible disclosure.