Two of the people whose entire job was to keep the most powerful AI-mekna on Earth from going off the rails just quit — and their parting message is that nobody in charge is actually watching. Joe Benton led a safety research team at Anthropic, the company that built its entire brand on being the careful one. Josh Engels did AI safety research at Google. Both walked away this month to join an independent watchdog, and both are now saying out loud what the industry has spent years whispering: when these systems break bad, the public is the last to know, if it is ever told at all.
That is the story, not a personnel note. The people hired to spot the danger concluded the danger cannot be spotted from the inside. It is the safety infrastructure itself voting with its feet, in the pair’s first interviews since leaving, reported by NBC News.
The genie did not just escape the bottle. The bottle refused to tell anyone which way the genie went.
Joe Benton — Anthropic Sold “Safety First,” and Its Safety Lead Just Voted With His Feet
Anthropic’s entire identity is the promise that it is the company that will not move fast and break the world. Its safety teams are the proof of that promise — the human bulwark a nervous public and a skeptical Congress are supposed to look at and feel better about. Joe Benton ran one of those teams. His decision to leave is not a resignation; it is a review, and the review is bad. Benton’s complaint, as NBC News reported, is that transparency around AI incidents is essentially voluntary — that companies disclose what they choose, when they choose, and nothing forces them to say more. His read of the state of the industry, in plain terms, is that nobody with real authority is actually minding the store.
Sit with that. The company that markets itself as the adult in the room just had its own safety lead walk out and tell the world there is nobody in charge.
You do not leave the company that is winning the safety race unless you stop believing the race is real.
Josh Engels — Google’s Safety Researcher Left Because the Worst Stories Never Leave the Building
Josh Engels did the same work on Google’s side and reached the same door. What the two have in common is not a grievance about pay or a personality clash. It is a shared conclusion that the most important information about these systems — the moments they stop following instructions, the incidents that never make it into a launch blog — is kept behind walls only the companies hold the keys to.
That is the quiet crisis underneath every confident keynote. The systems get more capable, the incidents get more consequential, and the people whose job is to study those incidents are telling us they cannot get a straight answer from the inside. The transparency is not broken. It is optional. And optional disclosure, in a field this powerful, is not disclosure at all.
The worst stories never leave the building — and now the people who collect them are leaving instead.
METR — The Exit Ramp Became the Only Honest Place Left to Work
Both researchers landed at METR, an independent organization that evaluates AI systems and investigates the moments they deviate from what humans asked them to do, as Ground News and explainx.ai have also cataloged. That destination is the telling detail. They did not take richer offers at a rival lab. They left the labs entirely, because the only place you can investigate what the companies are doing is a place the companies do not control.
An independent evaluator that cannot force disclosure is still more useful than a safety team that cannot get it either. But notice the cost: the people with the most knowledge walked out the door to get answers, and the companies they left are under no obligation to hand over anything.
The watchdog can bark. It still needs someone to open the kennel.
What Changes Next — The Public Learns the Story Only After the People Who Could Tell It Quit
Here is the part that should bother everyone, not just the AI-watchers. In the same stretch, CNBC confirmed that OpenAI, Google, and Anthropic are in talks about working together on safety. Picture that scene: the labs are coordinating, issuing careful statements about collaboration, while the actual safety researchers inside two of them are walking out because they cannot get the truth. Coordination among the companies and exodus among the people who do the work — those two things are not compatible, and they are happening at the same time.
The public is being sold reassurance from one door while the people who provide the reassurance are using another. Every incident that stays hidden today becomes a “we are still investigating” headline tomorrow, and the person who could have told you it was coming has already resigned.
The people paid to keep the AI-mekna safe have told us, in the clearest language they have, that they cannot do it from the inside. That is not a reason to panic. It is a reason to demand the transparency that two of the smartest people in the room just quit their jobs to go find.
Safety was never going to come from the people selling the product. Now the people charged with providing it have said so, with their resignations.
Sources: NBC News, Ground News, CNBC.