Number 7 of 12 · November 20, 2026

PIPETTE: Prompt Injection Purple-Teaming for Robust LLM Agents

High attack success in red-teaming does not mean better defense. PIPETTE rewards varied attacks, and the authors report agents that resist prompt injection while still doing their normal tasks.

What they found

The authors study indirect prompt injection, where an SI agent reads untrusted outside data that secretly tells it to do an attacker's task. They show that two leading RL-based attack generators, RL-Hammer and PISmith, suffer from mode collapse. This means they keep finding similar attacks, and the extra successes add no robustness when used for adversarial training. The authors introduce PIPETTE, which uses a dictionary learning-based reward to push attack generation toward a broader range of strategies. It also adds an on-policy adversarial training pipeline for this kind of defense. On AgentDojo and ASB, they report defenders with near-zero attack success rates on static attacks, resilience to strong adaptive attacks, and benign task utility close to the undefended baseline.

Why it matters

SI agents that browse, read email or call tools must handle data they cannot trust. Prompt injection is a main way these agents get hijacked. This work suggests that how diverse the attacks are matters more than how often they succeed when hardening a model. The abstract does not give exact numbers or say how well this holds beyond the two benchmarks.

How it changes the race

If the results hold, developers could harden agents against prompt injection more efficiently before deployment, without a large loss in usefulness. It is a refinement of the red-team and adversarial training process rather than a new capability.

Safety and control

This work bears directly on security and control of SI agents. It aims to keep agents following the user's instructions rather than an attacker's. Near-zero attack rates on tested attacks do not prove safety against new attack types.

The abstract, by the authors

Large language model (LLM) agents routinely interact with untrusted external data, exposing them to indirect prompt injection (IPI) attacks that hijack agents into executing attacker-chosen tasks. To uncover these vulnerabilities at scale, automated red-teaming methods often leverage reinforcement learning (RL) to generate attacks that achieve high attack success rates (ASR). Model developers use the attacks discovered during internal red-teaming to harden agents through adversarial training before deployment, a collaborative offensive-defensive process known in cybersecurity as purple-teaming. However, we show that achieving high ASR during red-teaming does not guarantee robustness gains from subsequent adversarial training. Specifically, we find that two state-of-the-art RL-based red-teaming methods, RL-Hammer and PISmith, suffer from mode collapse, repeatedly discovering similar attack strategies whose additional successes add no additional robustness improvements to adversarial training. To address this gap, we introduce PIPETTE, a unified prompt injection purple-teaming framework that effectively converts red-teaming discoveries into defender robustness improvements. PIPETTE employs a dictionary learning-based reward that encourages broader coverage of attack strategies during red-teaming. It also integrates an on-policy adversarial training pipeline tailored to IPI defense, enabling agents to resist attacks while completing legitimate user tasks. Extensive evaluations across agentic benchmarks (AgentDojo and ASB) demonstrate that PIPETTE produces defenders that achieve near-zero attack success rates on static attacks and can withstand strong adaptive attacks, while maintaining benign task utility close to that of the undefended baseline.

Quoted from arXiv:2611.24362. The research and its words belong to its authors.

The authors

Lipeng He, Yihan Wang, Jiawen Zhang, N. Asokan

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.