Automotive sensor reliability specs were built around known workloads: radar processing, camera ISPs, LIDAR point-cloud pipelines. The thermal budget, voltage stress, and hot-carrier injection models were calibrated for those patterns. Transformer inference is a structurally different workload -- higher sustained utilization, tighter memory bandwidth saturation, more frequent DVFS transitions -- and most deployed sensors were validated against the wrong profile.
The consequence is that aging predictions based on existing qualification data are optimistic. A sensor that passes ISO 26262 ASIL-D requirements at design-time may drift out of spec earlier than the vehicle lifetime targets assume, because the compute demand was underspecified. The report also flags a compounding security risk: an aging sensor in a predictably degraded state is a more tractable attack surface than one operating within tight nominal bounds.
The validation fix is not complicated in principle: run the actual inference workload during qualification, measure junction temperature under that load, and refit the aging model to the observed stress profile. In practice it means every T&M/ATE setup that certifies automotive perception hardware needs an AI inference workload generator in the test rack, running the production model, not a synthetic stimulus. That is an equipment gap most test labs have not closed yet. The labs that close it in the next 12 months write the qualification methodology the rest of the industry will copy.