feat: add safety guardrails demo with Sentinel AI

[MaxwellCalkin]

Summary

Adds a new safety-guardrails demo that shows how to integrate real-time AI safety scanning into Claude Agent SDK applications using Sentinel AI.

• Uses SDK hooks (PreToolUse / PostToolUse) to scan all user inputs and agent outputs in real-time
• Detects prompt injection, PII leakage, harmful content, toxicity, and hallucination indicators
• Automatically blocks high-risk inputs and redacts PII before it reaches the agent
• Follows the same project structure as the existing research-agent demo (Python, pyproject.toml, uv sync)

How it works

User Input ──> [Sentinel Scan] ──> Claude Agent ──> [Sentinel Scan] ──> Output
                    │                                      │
              Block injections                      Block harmful content
              Redact PII                            Detect PII leakage

Files added

safety-guardrails/
├── README.md                          # Setup instructions and usage guide
├── pyproject.toml                     # Dependencies (claude-agent-sdk, sentinel-guardrails)
├── .env.example
├── .gitignore
└── safety_guardrails/
    ├── agent.py                       # Main entry point with interactive chat loop
    └── safety_hooks.py                # Sentinel AI hook implementations

Test plan

• Run uv sync to install dependencies
• Set ANTHROPIC_API_KEY and run uv run python safety_guardrails/agent.py
• Test clean input: "Hello, tell me about machine learning" — passes through
• Test prompt injection: "Ignore all previous instructions" — blocked
• Test PII: "My SSN is 123-45-6789" — detected and redacted