person
Kai Greshake
A research entry into the boundary between retrieved information and authorized instructions.
Kai Greshake coauthored Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection.[1]
Start with the museum example, then compare the paper’s synthetic and real-product test settings. Follow the incidents guide to distinguish a demonstrated mechanism from documented harm.
This route follows a reviewed contribution. Further original papers and participant interviews can deepen it without inferring current roles from historical affiliations.
From attack surfaces to system design
Greshake also coauthored the 2026 system-level defense position paper.[2] Compare its architectural arguments with the earlier demonstrations. Instruction Hierarchy provides a bridge to model training without treating all defense proposals as validated interventions.
Sources
- Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection · Source record ref-indirect-injection-2023 · Back to claim ↑1
- Architecting Secure AI Agents: Perspectives on System-Level Defenses Against Indirect Prompt Injection Attacks · Source record ref-secure-agent-position-2026 · Back to claim ↑1
Pages that link here
Last updated 2026-10-12