Researchers introduce FLASH, a pulsed-light attack that exploits temporal HDR fusion in modern cameras by creating cross-exposure inconsistencies, causing darkening and visibility loss across eight camera platforms including smartphones and automotive systems. The attack requires no physical access or knowledge of fusion algorithms and successfully triggers low-visibility responses and severe image degradation in real-world scenarios like autonomous driving systems.
San Francisco tech workers at DrivingBench hooked ChatGPT and other large language models to a Toyota Corolla to test whether untrained AI chatbots could navigate a cone course in a parking lot. GPT-6 Astra successfully completed the course while other models failed, demonstrating that frontier LLMs can perform real-world driving tasks despite lacking specialized training.