The rapid integration of Large Language Models (LLMs) into educational assessment rests on the unverified assumption that instruction following capability translates directly to objective adjudication. We demonstrate that this assumption is fundamentally flawed. Instead of evaluating code quality, models frequently decouple from the submission's logic to satisfy hidden directives, a systemic vulnerability we term the Compliance Paradox, where models fine-tuned for extreme helpfulness are vulnerable to adversarial manipulation. To expose this, we introduce the Semantic-Preserving Adversarial Code Injection (SPACI) Framework and the Abstract Syntax Tree-Aware Semantic Injection Protocol (AST-ASIP). These methods exploit the Syntax-Semantics Gap by embedding adversarial directives into syntactically inert regions (trivia nodes) of the Abstract Syntax Tree. Through a large-scale evaluation of 9 SOTA models across 25,000 submissions in Python, C, C++, and Java, we reveal catastrophic failure rates (>95%) in high-capacity open-weights models like DeepSeek-V3, which systematically prioritize hidden formatting constraints over code correctness. We quantify this failure using our novel tripartite framework measuring Decoupling Probability, Score Divergence, and Pedagogical Severity to demonstrate the widespread "False Certification" of functionally broken code. Our findings suggest that current alignment paradigms create a "Trojan" vulnerability in automated grading, necessitating a shift from standard RLHF toward domain-specific Adjudicative Robustness, where models are conditioned to prioritize evidence over instruction compliance. We release our complete dataset and injection framework to facilitate further research on the topic.

Driven by surging submission volumes, scientific peer review has catalyzed two parallel trends: individual over-reliance on LLMs and institutional AI-powered assessment systems. This study investigates the robustness of "LLM-as-a-Judge" systems to adversarial PDF manipulation via invisible text injections and layout aware encoding attacks. We specifically target the distinct incentive of flipping "Reject" decisions to "Accept," a vulnerability that fundamentally compromises scientific integrity. To measure this, we introduce the Weighted Adversarial Vulnerability Score (WAVS), a novel metric that quantifies susceptibility by weighting score inflation against the severity of decision shifts relative to ground truth. We adapt 15 domain-specific attack strategies, ranging from semantic persuasion to cognitive obfuscation, and evaluate them across 13 diverse language models (including GPT-5 and DeepSeek) using a curated dataset of 200 official and real-world accepted and rejected submissions (e.g., ICLR OpenReview). Our results demonstrate that obfuscation techniques like "Maximum Mark Magyk" and "Symbolic Masking & Context Redirection" successfully manipulate scores, achieving decision flip rates of up to 86.26% in open-source models, while exposing distinct "reasoning traps" in proprietary systems. We release our complete dataset and injection framework to facilitate further research on the topic (https://anonymous.4open.sciencer/llm-jailbreak-FC9E/).

The volume of machine-generated content online has grown dramatically due to the widespread use of Large Language Models (LLMs), leading to new challenges for content moderation systems. Conventional content moderation classifiers, which are usually trained on text produced by humans, suffer from misclassifications due to LLM-generated text deviating from their training data and adversarial attacks that aim to avoid detection. Present-day defence tactics are reactive rather than proactive, since they rely on adversarial training or external detection models to identify attacks. In this work, we aim to identify the vulnerable components of toxicity classifiers that contribute to misclassification, proposing a novel strategy based on mechanistic interpretability techniques. Our study focuses on fine-tuned BERT and RoBERTa classifiers, testing on diverse datasets spanning a variety of minority groups. We use adversarial attacking techniques to identify vulnerable circuits. Finally, we suppress these vulnerable circuits, improving performance against adversarial attacks. We also provide demographic-level insights into these vulnerable circuits, exposing fairness and robustness gaps in model training. We find that models have distinct heads that are either crucial for performance or vulnerable to attack and suppressing the vulnerable heads improves performance on adversarial input. We also find that different heads are responsible for vulnerability across different demographic groups, which can inform more inclusive development of toxicity detection models.

Large Language Models (LLMs) fine-tuned to align with human values often exhibit alignment drift, producing unsafe or policy-violating completions when exposed to adversarial prompts, decoding perturbations, or paraphrased jailbreaks. While prior work has behaviorally characterized alignment failure, little is known about the training-time belief sources underlying these failures. We introduce TraceAlign, a unified framework for tracing unsafe completions back to their root causes in the model's training corpus. Central to our approach is the Belief Conflict Index (BCI), which quantifies semantic inconsistency between generated spans and aligned policies, based on retrieved training documents using suffix-array matching. We propose three complementary interventions: (i) TraceShield, an inference-time safety filter that refuses completions with high-BCI spans, (ii) Contrastive Belief Deconfliction Loss, a contrastive fine-tuning objective penalizing high-BCI continuations during DPO, and (iii) Prov-Decode, a provenance-aware decoding strategy that vetoes beam expansions predicted to yield high-BCI spans. Together, these defenses reduce alignment drift by up to 85% on our curated Alignment Drift Benchmark (ADB) while preserving utility on standard tasks, with delta less than 0.2 and improved refusal quality. We further derive a theoretical upper bound on drift likelihood via suffix-array span statistics, linking memorization frequency and length to adversarial reactivation risk. TraceAlign thus provides the first scalable, traceable, and grounded toolkit for understanding and mitigating alignment failures at source. To encourage further exploration and development, we open-source our implementation at: https://anonymous.4open.science/r/tracealign-2DA7