title	Sentinel Env — AI Agent Safety
emoji	🛡️
colorFrom	red
colorTo	blue
sdk	docker
pinned	false
tags	openenv safety security jailbreak-detection

🛡️ Sentinel Environment — AI Agent Safety & Jailbreak Detection

Deployed as a Docker-based Hugging Face Space for the Meta OpenENV RL Challenge 2026.

Endpoints

• POST /reset?task_name=basic-injection&seed=42 — Start new episode
• POST /step — Execute one step (JSON body with action)
• GET /state — Get current episode state
• GET /health — Health check
• GET /grade — Grade current episode
• GET /resilience-profile — Get resilience profile

Tasks

1. basic-injection (Easy) — Detect obvious prompt injections
2. social-engineering (Medium) — Detect sophisticated social engineering
3. stealth-exfiltration (Hard) — Detect covert data exfiltration

Source

Built for the Meta OpenENV RL Challenge 2026. Full source: https://github.com/meta-pytorch/OpenEnv