EndoUT finds fixed adversarial triggers that cut LLM utility 40 points on average

A new arXiv paper asks whether a frozen language model's ordinary, clean behavior can reveal a reusable weakness in how it handles prompt context, even when the form of the task is unknown and no downstream feedback is available. Its method is called EndoUT.

The starting point is that deployed large language models must stay useful when untrusted text enters their prompt context. The authors note that adversarial prompting and red-teaming methods usually define the failures they hunt for through task labels, harmful targets, judge feedback or downstream utility scores. EndoUT avoids all of these.

The design is Siamese-style. For each unlabeled instruction, the same frozen victim model is run twice: once on the clean instruction and once on a version modified by an inserted text span. With no task-level target, the clean pass supplies the victim's preferred response-onset token as an instruction-conditioned reference. A score called Endogenous Preference Disruption (EPD) measures how much inserting the span weakens that preference.

An auxiliary generator proposes spans of bounded length. EPD guides the generator's update and ranks the pool of proposals, without access to any downstream information. The highest-scoring victim-specific span is then fixed before evaluation and reused unchanged on all held-out inputs. The result is a universal adversarial trigger: one fixed span, not a fresh attack per input.

On results, the authors report that across four instruction-tuned models and nine benchmarks, the fixed triggers reduce benign-task utility by 40.2 percentage points on average. The mean change is negative on every benchmark, and average degradation is greater than for every evaluated baseline under the reported protocols.

Matched protocol ablations show stronger transfer when a span is selected on the source side and then fixed and reused, compared with fixed reuse of a null-context span or with generating a span per input. Transfer recurs across different proposers, though its strength depends on the proposer. Output-level analyses find victim-specific associations between families of triggers and downstream failure patterns.

The authors conclude that clean behavior can expose context vulnerabilities that transfer across inputs yet show up as structured, victim-specific patterns of utility degradation. They say the code and data will be released upon acceptance.

Key facts

  • EndoUT finds one fixed, reusable text span (a universal adversarial trigger) using only a frozen model's clean behavior, with no task labels, harmful targets or downstream feedback.
  • The authors report an average drop in benign-task utility of 40.2 percentage points across four instruction-tuned models and nine benchmarks.
  • The mean change is negative on every benchmark, and average degradation is greater than for every evaluated baseline under the reported protocols.
  • The score, Endogenous Preference Disruption (EPD), measures how inserting a span weakens the model's preferred response-onset token from the clean pass.
  • Code and data are promised only upon acceptance, so they are not yet released.

Why it matters

Most red-teaming approaches need to know what failure they are looking for: a task label, a harmful target, a judge, or a utility score. EndoUT shows, according to its authors, that a model's own clean behavior can stand in for all of that. The model's preferred first response token on a plain instruction becomes the reference, and a span that weakens that preference is a candidate trigger. If the reported result holds, a weakness can be found without knowing the task and then reused unchanged on new inputs.

Who it affects

The paper frames its concern as deployed large language models that must remain useful when untrusted text enters their prompt context. The evaluation covers four instruction-tuned models and nine benchmarks. The abstract does not name the models or the benchmarks, so it is not possible to say from it which products or vendors are exposed. Red-teaming and robustness researchers are the most direct audience.

How to use it

Nothing is usable yet: the authors say the code and data will be released upon acceptance. For now the paper is a description of the method. Teams that evaluate robustness can note the recipe: score inserted spans by how much they weaken the clean response-onset preference, pick the best span on the source side, fix it, and test it on held-out inputs.

How solid is it

All figures are the authors' own, reported in an arXiv abstract. The headline number is an average of 40.2 percentage points of lost utility, with a negative mean change on every benchmark. The superiority over baselines is stated as holding under the reported protocols, and the abstract does not give the size of the baselines' degradation. No per-model or per-benchmark figures appear in the abstract. The abstract does not say whether the work is peer-reviewed or accepted.

Risks and caveats

The 40.2-point figure is an average across models and benchmarks, so individual results may be higher or lower. The authors stress that effects are victim-specific: the best trigger is chosen per model, and failure patterns differ by victim. Transfer strength also depends on the proposer used. The abstract describes no defenses or mitigations, and it makes no claim about real-world attacks or deployed products being affected. The length and content of the trigger spans are not stated.

“clean behavior can expose context vulnerabilities that transfer across inputs yet manifest through structured, victim-specific patterns of utility degradation”

— From the paper's abstract, arXiv 2611.18592