Goodfire Launches Inside-Out AI Monitors to Catch Rogue Agents
This approach provides a significantly cheaper way to monitor AI agents for dangerous behaviors like hacking and weapons misuse during operation.
• Goodfire launched monitors that check internal AI model signals instead of reading all output text.
• In tests, monitoring 1 million exchanges cost roughly 185 dollars and caught 93 percent of malicious hacks.
• The monitors add less than 2 percent to response start times while operating on Baseten.
Read more
October 9, 2026 18
