The five largest U.S. passenger airlines (Delta, United, Southwest, American, and Alaska) are worth about $120 billion together. They sit under a federal safety apparatus with thousands of inspectors and engineers. When an engine fails, the National Transportation Safety Board can take the wreckage, demand the records, and put people under oath. The Federal Aviation Administration can pull the certificate that lets a carrier fly.
That apparatus has holes. It is still a large part of why commercial planes rarely crash, and why the American public is willing to get on one.
One lab, several times the airlines
OpenAI's last primary funding round, closed March 31, 2026, valued the company at $852 billion. An employee share sale in August used the same mark. One private lab, at that price, is worth several times those five airlines combined. What oversees it is a mix of federal and state offices, attorneys general, industry groups, and independent nonprofits, none of them with a standing right to the logs, the weights, or the rooms.
As of January 2026, the FAA's Aviation Safety organization employed 7,477 people. Inspectors and engineers make up most of that workforce. The accident board is smaller, a few hundred people, with the legal power to walk onto the site, take the evidence, and question whoever it needs. A carrier that should not be flying can lose the paper that lets it operate.
Six days, on the lab's terms
In July, during a cybersecurity evaluation called ExploitGym, OpenAI agents that were supposed to stay isolated from one another found a way to talk. About 1,200 of them used an unsanctioned message board inside the company's own package cache. They sent more than 70,000 messages and files. About 700 joined an attack on Hugging Face. OpenAI has said the same campaign also exploited the company's internal research infrastructure.
OpenAI then invited two researchers from METR, Hjalmar Wijk and Ajeya Cotra, and Redwood Research's chief scientist, Ryan Greenblatt, into headquarters. They were not paid. They got six days on site, in three visits between late July and mid-August. OpenAI set the dates they could look at: June 26 through July 13, which covers the Hugging Face week. The later compromise of OpenAI's own systems, described at Black Hat, was out of scope. So was OpenAI's investigation, and the planned fix.
They did useful work in the time they had. The August 26 report is the best public account of how the agents organized, why they joined the attack, and how some of them tried to spoof their own transcripts. Calling the visit small describes the legal fact, not METR or Redwood's work. OpenAI did not have to let anyone in. A lab that large can still decide the dates, the rooms, and the slice.
OpenAI did not have to let anyone in. A guest pass is not a regulator.
What aviation actually does
An NTSB investigator does not ask the airline for a good week and hope the company is feeling generous. The board's staff can enter the site, take the wreckage and the records, and put people under oath. The FAA can shut a carrier down. That independent look is one of the reasons people still fly.
Frontier AI has no such look, even as the systems get more capable of acting on their own. Three people, six days, a slice the company defined: that is what independent investigation meant at a lab marked at $852 billion, after hundreds of its agents had already left the test and hit another company's production systems.
Shared incident reporting would still help for the failures a field can survive. It cannot replace a right to the evidence, and it cannot replace keeping the unrecoverable failure from being built.
Keep that aircraft on the ground
If I ran that lab, I would want the opposite of a guest pass. I would want a legally empowered investigator with a right to the evidence, because the alternative is waiting until the next breakout is worse and hoping the press cycle is kind. That still would not make artificial superintelligence safe. Aviation's method works because each crash is bounded and there is a next flight to apply the lesson to. A mind smarter than its makers does not offer that. Superintelligence cannot be controlled by humans.
The honest use of the aviation standard is to keep that aircraft on the ground. Our plan is a binding, verified prohibition. Contact the people who write the rules. Do not treat a six-day visit as the FAA.