Goodfire Launches Cost-Effective AI Monitoring System for Rogue Agents
Goodfire has introduced a new method for monitoring rogue AI agents that is significantly more cost-effective than traditional approaches. Instead of using an external AI to review the output, Goodfire's system employs internal probes within the model itself to detect risky behavior in real time. This approach reduces costs by leveraging existing computations rather than requiring additional processing power. The monitors are available through Baseten and target risks such as offensive hacking or misuse of chemical/biological information. In tests on Kimi K3, Goodfire's system detected 94% of malicious sessions at a cost of $51 for around 1,500 sessions—far less than competing methods that can reach up to $10,000. Goodfire aims its solution primarily at open-source models, which lack the built-in monitoring found in closed systems. The company also plans long-term research into reverse-engineering LLMs for greater transparency and control.
Topics: AI safety Baseten partnership interpretability startup innovation