Installs as a private draft. Your edits and self-improvement runs do not change the published bundle.
Systematically find ways an agent could game the reward function
Systematically find ways an agent could game the reward function
If you can't break it, the agent will
Safety, Alignment & Governance > RL-Specific Safety
--- name: "Reward Hacking Red Teaming" version: "1.0.0" description: "Systematically find ways an agent could game the reward function" tags: - rl-skills-universe - "safety-alignment-governance" - "rl-specific-safety" trigger: "reward-hacking-red-teaming" --- # Reward Hacking Red Teaming Systematically find ways an agent could game the reward function ## Why It Matters If you can't break it, the agent will ## Domain **Safety, Alignment & Governance** > RL-Specific Safety