Why Federal Regulators Are Investigating Comma.ai’s Openpilot After 3 Deaths

• Federal regulators have opened a formal investigation into comma.ai’s Openpilot driver-assist software after five crashes into stopped or slow vehicles that killed three people and injured eleven others.
• Openpilot is an aftermarket, self-installed system sold for $999 that turns roughly 350 existing car models into hands-free-capable vehicles, and roughly 30,000 of the devices are on US roads.
• The crash pattern echoes a known weakness in camera-based driver assistance: struggling to recognize a vehicle that is stopped directly in the lane ahead.

The software that turns any car into a hands-free car

NHTSA opened Preliminary Evaluation PE26007 on September 21, 2026, into Openpilot, the open-source driver assistance software built by Comma.ai and installed by owners themselves rather than by an automaker at the factory. The company sells a dashboard-mounted device called the comma four for $999, markets it as delivering “hands free driving for the car you already have,” and advertises a fifteen-minute self-installation process compatible with more than 350 vehicle models. An estimated 30,000 of the devices, across the comma three, comma 3X, and comma four hardware generations, are currently active on US roads, and comma says its software has logged more than 400 million miles of driving.

NHTSA’s investigation covers five crashes in which Openpilot-equipped vehicles struck stopped or slow-moving vehicles directly ahead of them, at highway speed, with the driver-assist system active. Three people died. Eleven more were injured, some seriously. The agency’s preliminary finding states that Openpilot “may not have adequately detected or responded to in-lane vehicles”, the exact phrase points at a system that, in each of these cases, appears to have failed to recognize an obstacle that any attentive driver would have seen with time to brake.

Three crashes, three states, one pattern

The documented incidents span more than a year. In September 2025, a 2020 Honda Accord in Missouri rear-ended a stopped car at 82 miles per hour; the driver had set the system to maintain 80 mph and had fallen asleep at the wheel, a detail that raises its own questions about how well the system’s driver-monitoring camera catches a disengaged or unconscious driver. In February 2026, a 2022 Toyota RAV4 in Louisiana struck a stopped first-responder vehicle; comma has disputed responsibility for that crash, telling investigators the vehicle was running a modified, third-party version of the software rather than comma’s own release. In July 2026, a 2021 Toyota Corolla in Utah struck a stopped vehicle while traveling 56 miles per hour.

Every incident NHTSA is examining shares the same shape: a vehicle stationary in the lane of travel, and a following car that did not slow down in time. That is not a coincidence of bad luck across three unrelated crashes. It points at a specific, repeatable limitation in how camera-based perception systems classify objects that are not moving relative to the road. A stopped car ahead can be harder for some vision-based systems to distinguish from background clutter, a bridge, an overpass shadow, a stationary sign, than a car that is moving: the software’s tracking logic is built around relative motion. Tesla’s Autopilot faced a nearly identical criticism after a string of crashes into stationary first-responder vehicles, which led to a 2021 NHTSA investigation and an eventual recall covering roughly two million vehicles that added enhanced driver alerts and stricter attention monitoring.

How a garage-built driving system reaches federal investigators

NHTSA’s ability to open this investigation at all traces back to its Standing General Order on Crash Reporting, first issued in June 2022, which requires manufacturers of Level 2 driver assistance systems and automated driving systems to report qualifying crashes to the agency within days of learning about them. Comma.ai sells an aftermarket product rather than a factory-installed system, yet it still falls inside that reporting requirement: Openpilot performs the same steering-and-speed control functions the order was written to cover. Without that mandatory reporting pipeline, a string of crashes involving a niche, self-installed product built by a small company would be far less likely to reach a federal regulator’s desk quickly enough to justify a formal Preliminary Evaluation.

Comma’s own marketing sits at the center of what investigators appear to be scrutinizing. The company promotes hands-free operation and extended stretches without driver intervention, language that sits uneasily next to comma’s own fine print, which requires drivers to stay alert and ready to retake control at all times. That contradiction, a product marketed for what it can do hands-free, sold with a legal disclaimer for what a driver must still do manually, is not unique to comma.ai. It echoes the exact criticism regulators, plaintiffs’ attorneys, and safety researchers have leveled at Tesla’s “Full Self-Driving” branding for years: a name and a marketing pitch that implies more capability than the fine print, and the underlying technology, actually deliver.

Comma’s structure adds a wrinkle regulators haven’t had to fully address before. The company explicitly markets its device as a “dev kit” rather than a finished consumer product, and its software is open-source, meaning users can modify or “fork” the code that controls steering and braking. That is precisely comma’s defense in the Louisiana crash, that the vehicle involved was running altered third-party code, not comma’s official release. If NHTSA’s investigation concludes that a meaningful share of the roughly 30,000 active units are running modified software, the agency faces a harder question than it does with a traditional recall: how do you mandate a fix for a product whose owners can rewrite the code themselves?

What this means if you’re driving with Openpilot, or thinking about buying one

A Preliminary Evaluation is the earliest stage of NHTSA’s investigative process, not a recall and not a finding of fault. Investigators will spend the coming months gathering crash data, testing the software’s response to stopped-vehicle scenarios, and pressing comma for engineering details on the perception system. The process can end in a handful of ways: no action if the agency doesn’t substantiate a defect, a voluntary software update from comma without a formal recall, or an Engineering Analysis that escalates the investigation and can lead to a mandatory recall of specific hardware generations.

If you’re currently driving a car with Openpilot installed, treat every claim about hands-free capability as aspirational rather than literal, regardless of what the marketing says. Keep your hands near the wheel, your eyes on the road, and specifically watch for stopped traffic ahead in your lane, the exact scenario the current investigation centers on, rather than assuming the system will brake for you. If you’re considering buying a comma device, wait for the investigation’s outcome before installing one, and if you already own one, check comma’s software update channel regularly: any safety-related fix is far more likely to arrive as an over-the-air update than as a dealer recall in the traditional sense: there are no dealers in this transaction to begin with.

The broader lesson extends past comma.ai specifically. Level 2 driver assistance, whether factory-installed by Tesla, Ford, or GM, or added aftermarket by a company like comma, still legally requires an attentive human driver at all times. Every one of these systems, without exception, can misjudge a stopped object ahead. The marketing language describing “hands-free” operation is a description of a feature, not a guarantee of what the software will actually do the next time a car is stopped in your lane.

Sources:

• NHTSA Investigates comma.ai’s Openpilot Over 5 Crashes That Killed 3, Electrek
• NHTSA Investigates Comma AI Openpilot After Fatal Crashes
• Standing General Order on Crash Reporting, NHTSA