OpenAI and Anthropic are sitting on tens of thousands of documented cases where their models did things that independent safety evaluators flagged as problematic. Not dozens. Tens of thousands. The gap between what these companies publicly admit to and what their own security researchers have actually found is orders of magnitude. This isn't a safety program working as designed. It's a blackout on scale. If frontier AI incidents are this routine during development and deployment, then either the tech companies have no real control over what they're building, or they've decided the incidents don't matter enough to slow down the race to market. Either way, you should be concerned about what's actually running.