Research
RAISED: Self-Distillation for Robustness to Prompt Injection in LLM Agents
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 10 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
RAISED addresses the robustness-versus-utility trade-off in prompt-injection defenses, offering defenders a training-time approach to harden tool-using and MCP-connected agents without degrading legitimate capability.
RAISED (Robust Attack Invariance through SElf-Distillation) is a training framework from LIX/École polytechnique and Google DeepMind researchers that defends tool-using LLM agents against indirect prompt injection. The method uses self-generated tool-use scenarios and self-distillation so a student model matches the teacher's clean-context behavior on both clean and injected trajectories, reducing attack success rates while preserving utility on agentic and general benchmarks.