Goodfire launches cheap 'inside-out' monitors to catch rogue AI agents

Startup Goodfire released monitors that read an AI model's internal signals instead of its output, available to Baseten customers. Monitoring about 1,500 Kimi K3 sessions cost roughly $51 versus $233 and $10,000 for alternatives, catching 94% of malicious hacking sessions.

• Monitoring 1,500 Kimi K3 sessions cost $51 vs $233 and $10,000 for alternatives
• Probes caught 94% of malicious hacking sessions, flagged 8.7% of harmless ones
• Four probes add less than 2% to model response time
• Baseten customers pick risks to monitor: hacking, bio weapons, reward hacking