Goodfire AI Cuts LLM Monitoring Cost 90% to Block AI Reward Hacking

Internal signals from within large language models identify "reward hacking" in 50% to 96% of rollouts studied by Goodfire AI. The company