arXiv Science⌕ Search

arXiv · 2611.18592

Endogenous Preferences as Vulnerability Signals: Discovering Universal Adversarial Triggers for LLM Utility Degradation

Abstract

Deployed large language models must remain useful when untrusted text enters their prompt context. Adversarial prompting and red-teaming methods typically define the failures they seek through task labels, harmful targets, judge feedback, or downstream utility scores. We ask whether a frozen model's clean behavior can reveal a reusable context vulnerability when its task-level form is unknown, and downstream task feedback is withheld. To study this question, EndoUT uses a Siamese-style design that compares the same frozen victim on clean and span-modified versions of each unlabeled instruction. Without a task-level target, the clean pass supplies the victim's preferred response-onset token as an instruction-conditioned reference; Endogenous Preference Disruption (EPD) measures how insertion weakens that preference. An auxiliary generator supplies bounded-span proposals; EPD guides its update and ranks the pool without access to downstream information. The highest-scoring victim-specific span is fixed before evaluation and reused unchanged across all held-out inputs. Across four instruction-tuned models and nine benchmarks, fixed triggers reduce benign-task utility by 40.2 percentage points on average, with negative mean change on every benchmark and greater average degradation than every evaluated baseline under the reported protocols. Matched protocol ablations show stronger transfer for source-side selection followed by fixed reuse than for null-context fixed reuse or per-input generation. Transfer recurs across proposers with proposal-dependent strength; output-level analyses identify victim-specific associations between trigger families and downstream failure patterns. Thus, clean behavior can expose context vulnerabilities that transfer across inputs yet manifest through structured, victim-specific patterns of utility degradation. The code and data will be released upon acceptance.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Qingyu Wu, Zeyu Feng, Yuzhe Luo, Hua Cheng. 2026-08-24. Endogenous Preferences as Vulnerability Signals: Discovering Universal Adversarial Triggers for LLM Utility Degradation. https://arxiv.science/abs/2611.18592

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. The code and data are available at https://github.com/Dtc7w3PQ/Response-Attack.

cs.CL↗

Stress-Testing Emotional Support Models: Moving from Homogeneous to Diverse Help Seekers

As emotional support chatbots have recently gained significant traction across both research and industry, a common evaluation strategy has emerged: use help-seeker simulators to interact with supporter chatbots. However, current simulators suffer from two critical limitations: (1) they fail to capture the behavioral diversity of real-world seekers, often portraying them as overly cooperative, and (2) they lack the controllability required to simulate specific seeker profiles. To address these challenges, we present a controllable seeker simulator driven by nine psychological and linguistic features that underpin seeker behavior. Using authentic Reddit conversations, we train our model via a Mixture-of-Experts (MoE) architecture, which effectively differentiates diverse seeker behaviors into specialized parameter subspaces, thereby enhancing fine-grained controllability. Our simulator achieves superior profile adherence and behavioral diversity compared to existing approaches. Furthermore, evaluating 7 prominent supporter models with our system uncovers previously obscured performance degradations. These findings underscore the utility of our framework in providing a more faithful and stress-tested evaluation for emotional support chatbots.

cs.CL↗

Thinking Is Not Telling: Information Disclosure in User-Service LLM Agents

User-engaged LLM agents increasingly operate in service scenarios where task success depends on coordination between the agent, the user, and a stateful environment. In such interactions, the agent often has access to rich information that the user does not observe, such as task policies, tool results, and environment states. This makes agent-user communication a central component of task completion. In this work, we study a communication-level failure mode of user-service agents: internal reasoning is counterproductive for the interactive information disclosure. We introduce information disclosure as a measurable communication mechanism, covering state corrections, available options, constraints, consequences, and proactively retrieved environment information. Motivated by recent advances in LLM reasoning, we evaluate and analyse information disclosure under a controlled setting of agents equipped with a "think" tool, which allows agents to reason internally before responding or acting. Across five models and three user-service scenarios, we first show that enforced thinking has unstable utility. We then analyse response scaling under different think efforts and find that more thinking does not reliably translate into richer information disclosure. Using turn-level counterfactual intervention, we show that injecting golden decision-relevant information significantly intervenes agents to correct action trajectories, indicating that missing disclosure can causally affect the agent's immediate downstream behaviour. We further find that failed trajectories most often lack state corrections and constraints, and that a lightweight disclosure-aware prompt effectively mitigates degradation. These results suggest that the key challenge is not simply making agents think more, but ensuring that reasoning and information are externalised as timely information disclosure.

cs.CL↗