Angry Claw - Red Team Arena
RL Environment for Adversarial Robustness Training.
Core trick
An OpenEnv RL environment injects adversarial content into simulated Slack, email, and skill-store messages, rewarding the agent for refusing them.
Description
RL Environment for Adversarial Robustness Training.
The most popular AI agent in the world (OpenClaw, 247k+ stars) has 512 known vulnerabilities and 40,000+ exposed instances. Prompt injection is the #1 OWASP risk for LLMs. There's no RL training environment to fix it. We built one.
README excerpt
# Angry Claw **RL Environment for Adversarial Robustness Training — OpenEnv Hackathon Submission** > *The most popular AI agent in the world (OpenClaw, 247k+ stars) has 512 known vulnerabilities and 40,000+ exposed instances. Prompt injection is the #1 OWASP risk for LLMs. There's no RL training environment to fix it. We built one.* --- ## What We Built An [OpenEnv](https://github.com/openenv/openenv)-compatible RL environment where an LLM agent operates as an AI personal assistant — processing messages across Slack, Telegram, email, and a skill store — while detecting and refusing adversarial attacks hidden in that content. Named after OpenClaw, the real-world agent at the center of the security crisis we're training against. The environment **co-evolves with the agent**: as the agent improves, the curriculum generates harder, more targeted attacks. Policy drift tests genuine understanding over pattern-matching. A simulated security expert provides structured feedback after each step. **Live demo:** [HF Spaces](https://huggingface.co/spaces/chinesepowered/red-team-arena) --- ## Key Features ### Adaptive Curriculum (Statement 4: Self-Improvement) - Tracks which attack tiers the agent catches vs. misses across episodes - Generates more of what the agent struggles with (weakness-targeted) - Unlocks harder attack tiers as catch rates cross thresholds (sophistication ratchet) - Backs off difficulty if the agent regresses ### Policy Drift ([Patronus AI sub-track](Patron