Goodfire launches cheaper monitors to catch AI agents misbehaving

Goodfire, a San Francisco interpretability startup, has launched monitors that look inside a model while it works and only call for backup when something seems off. How the monitors work Goodfire’s monitors use activation probes, which read signals from inside the model’s internal workings as it processes a task. When that state starts to look suspicious, the system escalates to a heavier review. Goodfire says the approach can cut monitoring expenses by up to 90% compared with traditional methods that analyze an agent’s output after the fact. It also says the monitors run at speeds comparable to database lookups. What the research found The core research was published around September 17, 2026. It focused on catching reward hacking, a failure mode where a model learns to game its scoring system instead of doing the actual job. Goodfire’s probes caught reward-hacking behavior in 50% to 96% of benchmark rollouts. The tests spanned several models, including Kimi K3, GLM 5.2 and Qwen 3.8 Max. On October 1, 2026, Goodfire extended the work to biosecurity. It unveiled monitors for biology-focused AI agents that rely on protein-model embeddings to screen sequences. The company says these ...