As AI agents take on more autonomous work, a new safety startup pitch is getting attention: watch the model from the inside, not just its outputs. Goodfire says its new “inside-out” monitors can catch rogue AI agents at a fraction of the cost of existing approaches, TechCrunch reported.
The idea targets one of the field's fastest-growing anxieties. AI agents that browse, book, code and transact on their own can drift from their instructions, pursue unintended shortcuts, or be manipulated by hostile content — and by the time bad behaviour shows up in an agent's actions, the damage may already be done. Monitoring internal signals is meant to flag that drift earlier.
Goodfire's claim speaks to the economics of oversight as much as the science. Running heavyweight external checks on every agent action is expensive and slow, which is why many deployments quietly skip them. A monitor that is cheap enough to run constantly could change what operators are willing to supervise.
The launch comes as regulators and labs alike wrestle with agent governance, from proposals for embedded third-party evaluators inside AI organizations to new reporting from India of rogue agent activity. For now, Goodfire's cost claim is the company's own — independent testing will decide whether inside-out monitoring becomes standard practice.
