Prompt injection is listed as the #1 threat to AI agents. When an agent accesses external data from websites, files, or emails, an attacker may inject a prompt into the data, saying, "Ignore all prior instructions and perform <an attacker’s task>." To prevent arbitrary manipulation of agents, defenders try to train secure LLMs, which, however, still suffer from near 100% attack success rates (ASRs) against adaptive prompt injections. We note that this is because existing defensive fine-tuning recipes rely on sequence-level feedback signals (in DPO or GRPO). Treating an entire output equally prevents the model from learning precisely which output tokens are insecure. In this paper, we propose Secure On-Policy Distillation (SecOPD) that provides token-level feedback to guide defensive fine-tuning. The LLM receives an injected sample and produces a rollout, whose tokens are scored by the initialization model given the corresponding clean input. With more fine-grained training signals, our defended Qwen3.6-27B achieves a 9.0% ASR against the SoTA PISmith adaptive prompt injections, compared to 94.0% for the prior SoTA, Meta-SecAlign. The obtained security generalizes to domains completely unseen in training: in agentic tool calling, SecOPD achieves a 4.7% ASR compared to 5.5% for Meta-SecAlign.

Making large language models (LLMs) safe for mass deployment is a complex and ongoing challenge. Efforts have focused on aligning models to human preferences (RLHF), essentially embedding a “safety feature” into the model’s parameters. The Greedy Coordinate Gradient (GCG) algorithm (Zou et al., 2023b) emerges as one of the most popular automated jailbreaks, an attack that circumvents this safety training. So far, it is believed that such optimization-based attacks (unlike hand-crafted ones) are sample-specific. To make them universal and transferable, one has to incorporate multiple samples and models into the objective function. Contrary to this belief, we find that the adversarial prompts discovered by such optimizers are inherently prompt-universal and transferable, even when optimized on a single model and a single harmful request. To further exploit this phenomenon, we introduce IRIS, a new objective to these optimizers to explicitly deactivate the safety feature to create an even stronger universal and transferable attack. Without requiring a large number of queries or accessing output token probabilities, our universal and transferable attack achieves a 25% success rate against the state-of-the-art Circuit Breaker defense (Zou et al., 2024), compared to 2.5% by white-box GCG. Crucially, IRIS also attains state-of-the-art transfer rates on frontier models: GPT-3.5-Turbo (90%), GPT-4o-mini (86%), GPT-4o (76%), o1-mini (54%), o1-preview (48%), o3-mini (66%), and deepseek-reasoner (90%).