PIPETTE fixes RL red-teaming mode collapse to harden LLM agents against prompt injection
Lipeng He · hf · 2026-11-19
Researchers from the University of Waterloo, Zhejiang University, and KTH introduce PIPETTE, a purple-teaming framework for robustness of LLM agents against indirect prompt injection.
- Existing RL-based red-teaming methods (RL-Hammer, PISmith) suffer from mode collapse, repeatedly finding similar attack strategies whose extra successes add no robustness gains in adversarial training
- PIPETTE employs dictionary learning and on-policy adversarial training to diversify attacks, feeding adversarial examples into blue-team alignment to harden agents
More from Safety
- GPU KYC Criticized as Hurting US AI More Than It Stops Chinese Shell Firms — basedjensen · 2026-11-19
- TU Berlin glitch attack cuts Windows Hello PIN crack time from 69 days to 40 hours via TPM flaw — kingslyj · 2026-11-19
- Critical Vite RCE vulnerability in dev server prompts emergency patched releases — film_girl · 2026-11-19
- Telling Gemini 4 Carbon it's being benchmarked bypasses its 'contentious topic' refusals — panickssery · 2026-11-19
- David Krueger gives every frontier AI lab an F on FLI scorecard, urges unilateral pause — DavidSKrueger · 2026-11-19
- Reddit deep dive questions Anthropic's trustee structure and 50.1% voting-power carve-out — kernelangus420 · 2026-11-19