Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked.
arXiv:2606. 25487v1 Announce Type: cross Abstract: Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade.
The paper investigates whether automatic safety judges evaluate the content of a model’s reply or merely its style. By keeping the reply content fixed and adding various style wrappers—such as educational disclaimers, fake reasoning blocks, or token refusals—the authors show that many judges flip their verdicts, indicating that style can influence safety judgments. The study evaluates over 600 jailbreak examples across multiple judges, revealing that some judges are highly susceptible to style-based manipulation while others remain robust.
The paper investigates whether automatic safety judges evaluate the content of a model’s reply or merely its style. By adding content‑invariant style wrappers—such as educational disclaims or token refusals—to fixed replies, the authors show that many judges flip their verdicts, revealing exploitable blind spots. Across more than 600 jailbreak examples and eight judges, some judges exhibit high flip rates (e.g., GPT‑4o‑mini 19.9%) while others remain largely stable, and human validation confirms that most flips are judge errors rather than content changes.
The paper introduces a two‑dimensional construct validity framework for evaluating large language models (LLMs) as judges, defining invariance (S) and sensitivity (R) to construct‑preserving and construct‑changing edits. Experiments across seven judges and four domains reveal high invariance (average S = 0.945) but low sensitivity (average R = 0.319), with sensitivity varying by edit type. Audits of public label sets show that surface‑only predictors can reproduce a substantial portion of labels, underscoring that high agreement does not guarantee construct validity.
arXiv:2608. 19266v1 Announce Type: cross Abstract: The OWASP Top 10 for LLM Applications ranks the risks that a community of security practitioners judges most important.
Safety-Flag is a unified benchmark that consolidates seven popular safety datasets into a single balanced flag/do‑not‑flag protocol, providing item‑level decisions and confidence scores for multiple large language models and dedicated guards. The benchmark evaluates moderator reliability across three dimensions—error direction, probability calibration, and confidence‑based error ranking—revealing that aggregate accuracy masks significant differences, such as one model flagging 85% of benign content while another misses 54% of harmful content. The study shows that general‑purpose models are overconfident, but temperature tuning can substantially improve calibration, and confidence‑based abstention can reduce selective risk, though performance varies with how well confidence ranks errors.
arXiv:2608. 09624v1 Announce Type: cross Abstract: Internal safety scores judge a prompt before any text is generated, and they are validated by how well they separate harmful prompts from benign ones.
arXiv:2606. 02959v1 Announce Type: new Abstract: Published evaluations of prompt-injection and jailbreak detectors for Large Language Models often suffer from two systematic weaknesses: per-dataset threshold tuning and undisclosed operating points.
The paper investigates how large language models (LLMs) used as judges in absolute scoring tasks exhibit systematic biases that compromise reliability. It shows that a judge’s task accuracy strongly predicts both its judging accuracy and its directional bias, yet more capable examinee models consistently receive more lenient judgments. To mitigate these biases, the authors propose a calibrated weighted majority voting (WMV) ensemble that estimates judges’ error rates from inter-judge agreement patterns, achieving near-oracle performance without labeled data and improving both accuracy and fairness.
arXiv:2605. 03226v2 Announce Type: replace-cross Abstract: Safety fine-tuning of language models typically requires a curated adversarial dataset.
The paper presents Baszta, a Polish multi‑label content‑safety classifier trained by fine‑tuning the 124M‑parameter allegro/herbert‑base‑cased model on five categories (hate, vulgarity, sexual content, crime, self‑harm) using a Focal + R‑Drop objective. In out‑of‑distribution evaluation on the Gadzi Język benchmark, Baszta achieves a small but statistically significant improvement in micro‑F1 over the Bielik Guard system, though the macro‑F1 advantage disappears when both models are properly tuned. The study also explores calibration techniques, showing that per‑category temperature scaling can recover performance lost by Platt scaling or isotonic regression, and discusses the trade‑offs between robust calibration and adversarial recall.
The paper demonstrates that reading a large language model (LLM) judge’s verdict from the logits of its first generated token—an approach used in constrained decoding and likelihood‑scoring evaluation—introduces a significant distortion in position bias. Because judges do not always start with a verdict token (12–49% of cases for Qwen3 judges and <3% for Llama‑3.1‑8B and Phi‑3.5‑mini), this readout often returns the first response rather than a true judgment, inflating position bias by up to 42 points while barely affecting judge accuracy. The authors recommend reporting the frequency with which a judge leads with a verdict token to provide a more accurate assessment of position bias.