Activation monitors from Goodfire identify "reward hacking" in AI models in real time by analyzing internal activations rather than output. These probes detected cheating or the gaming of metrics in 50% to 96% of rollouts studied across agentic benchmarks, identifying an internal signal associated with terms such as "sneak" and "illicit."

Implementation of these probes on the Kimi K3 model reduced monitoring costs by 90% with approximately a 1% decrease in precision compared to LLM judges. This efficiency makes monitoring feasible at scale and allows developers to identify broken environments, discover new failure modes, and stop reward hacking in progress while catching actions that appear innocuous to standard monitors.