Fatal Crash Occurred on a FrogPilot Fork of openpilot — Now NHTSA Is Investigating comma.ai
The PE26007 investigation, opened on September 21, 2026, means the question is no longer only about three deaths and 11 people seriously injured. It touches an unresolved principle: Can a company be held responsible for crashes that occur on third-party variants of its open-source code — and is it even obligated to report under the regime that requires it?
What is new
The National Highway Traffic Safety Administration (NHTSA) and its Office of Defects Investigation (ODI) opened a Preliminary Evaluation on September 21 under case number PE26007, as Electrek reported with reference to the investigation. It covers comma's three most recent hardware generations — comma three, comma 3X and comma four — as well as openpilot itself, the open-source software that runs on them (Electrek). The scope of the case was reported on September 23–24 by, among others, Electrek, Ars Technica and Yahoo/Business Insider.
The announcement states, in translation: ODI is opening this preliminary investigation to further examine incidents involving the comma devices and openpilot, particularly crashes involving slower or stationary vehicles or in-lane hazards. The investigation will gather data to verify system status in each individual incident, and evaluate system performance and the safety features intended to ensure the safe operation of vehicles equipped with comma devices (Yahoo, citing NHTSA ODI).
The crash record
The investigation builds on five collisions in which vehicles fitted with comma equipment allegedly struck stationary or slow-moving vehicles in the same lane. According to the initial data, the crashes resulted in three deaths and 11 serious injuries — figures given only as totals, not as a breakdown across the five individual cases (Yahoo, citing NHTSA ODI).
NHTSA writes in its review, in translation: ODI has confirmed that the systems were powered on or engaged in several of the crashes, and initial data ODI has reviewed suggests that the comma system may not necessarily have detected or responded adequately to the vehicles in the lane (Business Insider via Yahoo). The wording is deliberately cautious: NHTSA describes the findings as initial data that "may not" hold, not as an established causal link.
One of the crashes appears to match a February incident in Ascension Parish, Louisiana, in which a Toyota RAV4 struck an unoccupied police car that was stopped in a lane with its emergency lights on. Two passengers in the Toyota died. According to state police, the driving software was not identified in the police report (Business Insider via Yahoo).
comma's own position
comma has not commented publicly on the case in the source material. What exists of the company's position comes from its filings under NHTSA's Standing General Order — the regime requiring manufacturers of vehicles and certain driving systems to report serious incidents.
In the report on the fatal Louisiana crash, comma writes that the device, in translation, "appears to have been running third-party non-comma software" — in other words, a fork of openpilot (Electrek, citing comma's filing). The company also says this limits its ability to investigate what went wrong, but that it has provided the data NHTSA requested (Crypto Briefing).
Every one of comma's crash reports opens with the same disclaimer: that the information is provided without comma agreeing with NHTSA's interpretations, or conceding that comma is obligated to respond to the Standing General Order (Electrek). In other words: comma shares data, but refuses to concede that the company is subject to the reporting obligation of the very regime the data is filed under.
The fork question
At least two of the five crashes involved forks of the main build of openpilot, according to Ars Technica. The fatal crash ran on a version called FrogPilot (Ars Technica).
NHTSA is apparently preparing to meet this. According to Electrek, with reference to the investigation's basis, ODI will evaluate modifications to openpilot and assess components shared between openpilot and the forks. Electrek notes that popular community forks such as sunnypilot and FrogPilot build on the same driving model and the same codebase (Electrek).
This is what makes the case a test of principle: openpilot is open source, and other developers can copy the codebase and make their own changes. If a company can limit its liability by pointing out that the crashes occurred on variants that are not comma's, the question becomes whether NHTSA can pull in the same codebase and argue that because the code is largely identical, so is the responsibility.
Scale and counterpoint
comma itself states that more than 30,000 drivers use openpilot, with over 402 million miles (650 million km) driven. The comma four, according to the company, adds lane-keeping, adaptive cruise control and a cloud-based dashcam function to more than 325 car models from 27 brands (Ars Technica).
One of the five crashes also illustrates that the driver's role is part of the picture. In September 2025, a 2020 Honda Accord with openpilot verified as engaged collided with a stationary car on a Missouri highway, at 82 mph (132 km/h). According to Honda, citing the police report and a lawsuit, the driver had set cruise control to 80 mph and then fell asleep (Electrek). The same crash was also reported in by Honda itself — the vehicle manufacturer.
What remains unresolved
A Preliminary Evaluation is an initial inquiry, not a conclusion. It can lead to further analyses, an engineering analysis or a recall — or it can be closed. What NHTSA will land on is not known from the source material.
It is also unresolved how the allocation of responsibility between comma and the developers of forks like FrogPilot and sunnypilot will actually be handled, and how NHTSA will in practice weigh shared components against actual modifications. A civil lawsuit from Kevin McGinnis, filed on January 23, 2026 and connected to one of the incidents, was dismissed without a verdict in February 2026 — "without prejudice" — meaning it can in principle be refiled (Crypto Briefing). comma has not responded to questions in the source material.
The fact that one of the five crash reports (ID 30351-1807) is, according to Electrek, not in the released Standing General Order data means the details of that individual crash cannot be verified from available sources. The sources describe the crashes and their consequences as initial data, not as established fact. That is precisely the question the investigation is now to clarify.