Google details Gemini's layered defense strategy against indirect prompt injection
Mitigating prompt injection attacks with a layered defense strategy
AI summary
Google's GenAI security team introduced Gemini's layered defense strategy against indirect prompt injection, covering all stages of the prompt lifecycle. The strategy includes adversarial data training for Gemini 2.5 models to enhance model resilience, plus five built-in defenses: a prompt injection content classifier, strengthened safety thinking, Markdown sanitization and suspicious URL blocking, a user confirmation framework, and security mitigation notifications for end users.
Why it was selected
Google has released details of Gemini's layered defenses against indirect prompt injection, and teams deploying generative AI products can map them against its five protection layers.