The investigation
The US National Highway Traffic Safety Administration has opened a preliminary evaluation, numbered PE26007, into Comma.ai’s aftermarket driver-assistance hardware. The agency’s Office of Defects Investigation says it is looking at five crashes. Two were fatal and killed three people; across four of the crashes as many as eleven people were injured, some seriously.
The ODI states the common thread plainly: the devices failed to detect or respond to slow or stopped vehicles in the same road lane.
Comma.ai sells a device that retrofits hands-off lane-keeping and adaptive cruise control to ordinary cars, doing roughly what Tesla Autopilot and GM Super Cruise do on the vehicles that ship with them. Its software, openpilot, is open source.
The Louisiana crash
One of the two fatal crashes happened in February 2026 in Ascension Parish, Louisiana. A 2022 Toyota RAV4 struck a stationary police vehicle. Louisiana State Police said the car hit a fully marked Gonzales Police Department unit that was unoccupied, positioned in the left lane, with its emergency lights on. Two rear-seat passengers in the Toyota died.
The RAV4 was not running openpilot. It was running FrogPilot, a third-party fork of it.
That is the part of this investigation with no clear precedent. The ODI says preliminary findings indicate some of the crashes may have involved forked versions of the software, and a stated part of its work is establishing what role those modifications played. Regulators have investigated manufacturer driver-assistance systems before. They have not had to apportion responsibility between a company’s software and a community modification of it that the company did not ship.
What Comma already says
Comma.ai’s own website acknowledges the failure mode. It states that openpilot’s adaptive cruise control performance “can be affected by vehicles in the same lane that are not moving”, and that the system’s braking is “limited” for “abrupt braking maneuvers”. Comma.ai and its founder George Hotz could not be reached for comment on the investigation.
Stationary obstacles are a known weakness of radar-and-camera cruise control generally, not a quirk of this product, and disclosing it in the documentation is more than some vendors manage. Whether a documented limitation is a defect is exactly the question a preliminary evaluation exists to answer.
A preliminary evaluation is the first formal step, not a finding. It can close with no action, or escalate to an engineering analysis and from there to a recall request. What makes PE26007 worth following is the fork question. If the agency concludes that a manufacturer carries some responsibility for what third parties build on its released code, that reasoning will not stay inside the automotive sector.
