Google's AI Safety Guards Break Basic Search, Revealing a Hard Trade-off

Google's AI Overviews now misread words like "disregard" as commands to the system rather than dictionary queries, returning blank results. The problem stems from filters designed to block prompt injection attacks—where hidden instructions exploit AI systems. It exposes a central challenge: stronger defenses against manipulation degrade legitimate use. As AI spreads into consumer products, similar tensions between security and functionality will become routine.