AI safety startup Goodfire is selling tools that look inside AI models. Its probes cost 51 dollars for 1,500 sessions.

The probes work from the inside out. They track how a model forms its answers step by step.

The company claims 94 percent detection on the Kimi K3 model. It says the tools catch risky behaviour early.

AI safety has become big business. Regulators want proof that models are safe before release.

Goodfire's approach differs from outside testing. It reads the model's internal signals.

The startup raised fresh funding this year. It plans to support more models.

Experts say inside-out tools are promising. But they warn no tool catches every risk.

The probes plug into existing AI systems. No retraining is needed, the firm says. Customers include labs and regulators.

Pricing starts low to attract small teams. Enterprise plans cost more. The firm says demand has grown fast this year.

The company says more model support is coming. Independent tests of its claims are pending.