Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the cost

Most AI monitors use a second model to read everything an agent writes, which gets slow and expensive. Goodfire's monitors instead use small detectors that read the model's internal signals as it works and send only suspicious cases to a separate model for a closer look. In tests they caught 93 percent of malicious hacking sessions with almost no added delay.

Summary written with AI by HowTechMoves, based on reporting by TechCrunch.