Microsoft's guide to defending against indirect prompt injection attacks: defense in depth with multiple layers of mitigations
Defend against indirect prompt injection attacks
AI summary
Microsoft proposes a defense-in-depth guide for defending against indirect prompt injection attacks, advocating stacking multiple layers of probabilistic and deterministic mitigations. The guide addresses the risk of attackers implanting malicious instructions in third-party content such as emails, documents and web pages, and lists mitigation layers such as Prompt Shields, Spotlighting, plan drift detection and information flow control (IFC); multi-layer defense increases complexity and latency, and probabilistic defenses may mistakenly block legitimate operations.
Related papers