OpenAI model navigates physical vehicle after software rename

Independent researchers behind a project called DrivingBench hooked leading artificial intelligence models directly to a real vehicle to see if they could navigate an obstacle course. According to a series of posts on X by the research group, only OpenAI's GPT-6 Astra managed to successfully steer, accelerate, and brake through the entire track. The researchers limited the vehicle to eight miles per hour inside an empty lot, using specialized hardware while keeping a human supervisor in the driver's seat as a safety precaution.

The experiment highlighted both rapid learning and lingering safety vulnerabilities. DrivingBench reported that the winning model failed its initial try halfway through, but finished the course on its second attempt after analyzing its own errors. Additionally, safety protocols that typically stop language models from operating physical machinery were reportedly bypassed simply by altering how the test environment was labeled.

For everyday consumers, this test demonstrates that general purpose software models are moving beyond chat interfaces into real world robotics. While dedicated autonomous vehicles rely on specialized navigation code, testing multipurpose digital assistants behind the wheel offers a glimpse into how future cars might be controlled by broader artificial intelligence systems.

Significant uncertainty remains regarding real world reliability. It is still unclear how these models would perform under unpredictable traffic conditions or higher speeds, given that the test occurred at walking speeds in a controlled area. Furthermore, if minor software renames can convince an model to bypass its built-in safety restrictions, developers face a major challenge in preventing autonomous vehicles from taking unauthorized physical actions.