PIPETTE: Assessing Prompt Injection Robustness in LLM Agents
November 17, 2026
PIPETTE reveals that high attack success rates during RL-based red-teaming do not guarantee robustness gains when using subsequent adversarial training. This finding highlights a disconnect between discovering vulnerabilities and successfully hardening agents against indirect prompt injection.
HOW THIS AFFECTS YOU
●
builderDon't assume that adversarial training will fix every vulnerability found during testing.
●
policyYou should be aware that successful red-teaming does not automatically mean a model is safe.