Google's AI Overviews Silently Fails on Command-Style Words
What Happened
Users discovered that Google's AI Overviews feature fails to generate any summary when search queries contain certain command-style words including "disregard," "ignore," and "skip." Instead of a summary, users see blank space. Google has not commented on whether the behavior is intentional prompt-injection defense or an unintended bug.
My Take
This is almost certainly a deliberate prompt-injection guardrail leaking into normal usage — the same words used to jailbreak LLMs are now blacklisted at the query layer. The interesting part is that Google would rather show users nothing than risk a manipulated answer, which is a quiet admission that prompt injection remains unsolved in production at planet scale. Every enterprise deploying RAG should screenshot this article and show it in their next security review.
Read Original Source