Large language models (LLMs) remain vulnerable to jailbreak attacks that elicit structured, credible misinformation. Existing benchmarks primarily track binary refusals, often overlooking the severity and exposure risk of the misinformation payload itself. In this work, we introduce Misinfo-Jailbreak Bench, a comprehensive framework that systematically maps diverse attack strategies to a claim-level scoring pipeline for precise risk quantification. Building on this diagnostic foundation, we propose Selective LoRA Ethical Surgery, a parameter-efficient defense mechanism. Instead of broad parameter updates, our approach identifies safety-relevant subspaces via contrastive representations to confine LoRA edits and employs inference-time gating to preserve benign capabilities. Experimental results on state-of-the-art models (e.g., Qwen2.5, Llama-4) demonstrate that our method reduces high-severity risk by 60-70% while limiting general capability degradation to less than 1%, offering a precise and minimally invasive paradigm for post-hoc alignment.

Multimodal LLMs are increasingly used with multi-image, interleaved image-text inputs, yet hallucination evaluation remains largely single-image. We study target-binding hallucination: a model answers as if an object appears in the queried target image when it appears only elsewhere. We introduce a position-controlled COCO haystack with POPE-style object-existence queries and split negatives into clean cases (object absent everywhere) and distractor-present cases (absent from the target but present elsewhere). On InternVL2-8B at N = 10, clean false positives remain near floor (1.15%), whereas distractor-present cases reach 10.55% and show a position-sensitive profile; the same regime-level separation appears across three additional architectures and in an object-balanced PASCAL VOC2012 check. Channel and paired-removal controls are consistent with a visual-channel-dominant source-misattribution interpretation, and a related MIHBench Existence re-audit finds that over 98% of false positives concentrate in partial-present negatives. For object-existence diagnostics and other source-sensitive multi-image evaluations with auditable negative evidence, we recommend reporting clean and evidence-supported negatives separately whenever non-target images may contain supporting evidence.

We present DADS, a novel Data Augmentation technique for low-resource Dialogue Summarization. Our method generates synthetic examples by replacing sections of text from both the input dialogue and summary while preserving the augmented summary to correspond to a viable summary for the augmented dialogue. We utilize pretrained language models that produce highly likely dialogue alternatives while still being free to generate diverse alternatives. We applied our data augmentation method to the SAMSum dataset in low resource scenarios, mimicking real world problems such as chat, thread, and meeting summarization where large scale supervised datasets with human-written summaries are scarce. Through both automatic and human evaluations, we show that DADS shows strong improvements for low resource scenarios while generating topically diverse summaries without introducing additional hallucinations to the summaries.