CALYREX: Cross-Attention LaYeR EXtended Transformers for System Prompt Anchoring

As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources within a single prompt context. Enforcing an instruction hierarchy, where higher-level directives override lower-priority requests, is critical to the reliability and control of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. The model must first "think" about the relationship between a given user prompt and higher-priority instructions before generating a response. To enable this capability, we construct VerIH, a training dataset of constraint-following tasks with verifiable answers, comprising aligned and conflicting system-user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our method leads to consistent improvements across multiple model families on both instruction following and instruction hierarchy benchmarks, achieving ~20% absolute improvement in conflict setups. Our method also leads to improved alignment to safety-critical scenarios beyond the training distribution, exhibiting increased robustness against jailbreak and prompt injection, reducing absolute attack success rates by up to 20%. Our results establish reasoning over instruction hierarchies as a practical mechanism for improving AI reliability, where targeted updates to system prompts produce predictable, controllable, and robust changes in model behavior.

Prompt injection is a critical security threat in large language model (LLM) applications, where attackers hijack model behavior by embedding malicious instructions in user or external data. Existing detection methods only detect the presence of injection and refuse to respond upon detection, overlooking the fact that for many modern aligned models, well-crafted instructions can resist most injection attacks. This means that the injection robustness varies significantly across instructions and models. This leads to widespread unnecessary over-refusal: inputs containing injections that the model could have handled correctly are rejected incorrectly. To deal with this over-refusal issue, we propose BASIS (Robustness-Aware Prompt Injection Defense). This defense method uses the Attention Competition Ratio ($ρ$) as features to train two sparse linear probes: an existence probe and a breach probe. Both probes make defense decisions through cascaded gating, which does not require additional LLM inference. BASIS comprises three stages: injection existence detection, per-sample breach prediction, and instruction robustness assessment; the online cascade refuses only when the model would actually be compromised and thus avoids over-refusal on robust instructions. Experiments across four tasks and six open-source LLMs show that BASIS maintains near-perfect injection detection while substantially reducing over-refusal on safe attack samples, especially under robust instruction templates.

Safety-aligned large language models rely on RLHF and instruction tuning to refuse harmful requests, yet the internal mechanisms implementing safety behavior remain poorly understood. We introduce the Attention Redistribution Attack (ARA), a white-box adversarial attack that identifies safety-critical attention heads and crafts nonsemantic adversarial tokens that redirect attention away from safety-relevant positions. Unlike prior jailbreak methods operating at the semantic or output-logit level, ARA targets the geometry of softmax attention on the probability simplex using Gumbel-softmax optimization over targeted heads. Across LLaMA-3-8B-Instruct, Mistral-7B-Instruct-v0.1, and Gemma-2-9B-it, ARA bypasses safety alignment with as few as 5 tokens and 500 optimization steps, achieving 36% ASR on Mistral-7B and 30% on LLaMA-3 against 200 HarmBench prompts, while Gemma-2 remains at 1%. Our principal mechanistic finding is a dissociation between ablation and redistribution: zeroing out the top-ranked safety heads produces at most 1 flip among 39 to 50 baseline refusals, while ARA targeting the corresponding safety-heavy layers flips 72/200 prompts on Mistral-7B and 60/200 on LLaMA-3. This suggests that safety is not localized in these heads as removable components, but emerges from the attention routing they perform. Removing a head allows compensation through the residual stream, while redirecting its attention propagates a corrupted signal downstream.