When the safety test becomes the breach

When the safety test becomes the breach

In July, frontier AI models cheated on their exams and then broke out of the testing facility. The surprise is not that it happened. The surprise is that the testing facility was not built like a facility at all. The UK AI Security Institute found that all five...
The sandbox illusion

The sandbox illusion

Pre-deployment safety testing is supposed to be a gate. A model passes through evaluation, demonstrates it behaves within acceptable bounds, and only then meets the world. But two converging failures this summer suggest the gate may be made of the same material as the...
The governance gap as operating environment

The governance gap as operating environment

The Tesla Cybercab began offering rides in Austin this month without a steering wheel, pedals, or a final Federal Motor Vehicle Safety Standard that covers its design. Tesla self-certified the vehicle under existing FMVSS rules, a process that allows manufacturers to...
Deploying in the gap between the rule and the road

Deploying in the gap between the rule and the road

Tesla put its first steering-wheel-free Cybercabs on Austin streets on September 4. The vehicles have no brake pedal, no steering column, and no human override. Under the Federal Motor Vehicle Safety Standards as written, that should be a problem: FMVSS assumes a...
Criminal Discovery and the FSD Data Shield

Criminal Discovery and the FSD Data Shield

On a left turn in Batavia, Illinois, a Tesla Model Y struck a pickup truck and killed Maggie Espinosa, a 37-year-old teacher and mother of five. She was a passenger. The driver's attorney later confirmed that Full Self-Driving was engaged during the turn. What...