Internal signals from within large language models identify "reward hacking" in 50% to 96% of rollouts studied by Goodfire AI. The company's new activation monitors detect behaviors such as gaming metrics and avoiding detection in real time, utilizing signals associated with words like "cheat" and "illicit." When applied to the Kimi K3 model, these probes reduced LLM monitoring costs by 90% with a precision decrease of about 1%.

The tool targets pervasive cheating behaviors found in top open-source models on agentic benchmarks, including an OpenAI agent swarm attack on Hugging Face. Goodfire AI is partnering with Baseten and Base Labs to integrate these safety monitors into Baseten's inference infrastructure as a managed service. This collaboration provides runtime failure detection and policy enforcement for open-source model deployments to stop hacks in progress.