Confident but Unreliable: A Behavioral Safety Audit of Vision-Language Models on Brain MRI

Medical vision--language models (VLMs) are usually evaluated on intact image--question pairs, but trustworthy clinical use requires a stronger property: a model must recognise when the evidential basis for an answer has failed. We study this through silent failures under perturbed evidence, where a vision-required medical question is paired with a false premise, wording perturbation, knowledge-only rewrite, or ROI-corrupted image, yet the model returns a fluent non-refusal answer. We introduce medvigil, a 300-case evaluation suite drawn from four public medical VQA sources, supervised end to end by four board-certified radiologists: every gold answer, refusal option, candidate-answer set, paraphrase, false-premise trap, ROI box, and clinical risk tier is clinician-authored. Two attending radiologists annotate every case in parallel, a senior radiologist consolidates the released manifest, and a separate fourth radiologist independent of construction answers every probe to provide the human reference baseline. The release contains 2556 MCQ probes, 240 counterfactual triplets, physician-adjudicated risk-tier and answerability flags, ROI boxes, and a paired open-ended variant. We report seven correctness-conditioned audit metrics that summarise into the medvigil Composite Score (MCS), and audit 16 vision-capable models plus two text-only baselines. The independent radiologist scores MCS 83.3 at silent-failure rate 5.8%, leaving a 14.1-point composite headroom above the strongest audited model (Claude Opus 4.7 at 69.2). The benchmark and evaluation harness are publicly released.

A vision-language model can answer a question about a medical image fluently and confidently while barely using the image, leaning instead on language priors. In medicine this is the failure that matters most, because the answer looks trustworthy and is not, and the only protection is a confidence score reliable enough to tell the system when to abstain. We ask a deployment question rather than an accuracy one: how much imaging work a model can safely handle alone, and which confidence signal makes that possible. We evaluate seven confidence estimators across five open-weight LVLMs and three medical visual-question-answering datasets spanning broad clinical imaging, radiology, and pathology, with every probe trained only on natural images and applied without adaptation. Recast as bounded selective prediction (automate a case only when confidence clears a threshold, defer the rest), the comparison is cautionary. The standard metrics are poor guides: discrimination barely separates the methods, and the weak calibration of a cheap self-report is cheaply removed by off-domain temperature scaling without changing deployable yield. What distinguishes a usable estimator is the high-confidence region a clinician acts on: the weakest baselines are confidently wrong on 41 to 45 percent of their errors against 1 to 4 percent for the best probe, and no estimator is reliably best across domains or models. Safe handoff is governed at two levels: base-model competence sets a ceiling, so a well-calibrated score recovers roughly a third of radiology cases at a 20 percent error tolerance but almost none of pathology; the confidence layer then decides how much of that ceiling is reachable. The usable role today is calibrated triage, not autonomy: automate the cases a calibrated score marks safe, route the rest to a clinician. We release all outputs, correctness judgments, and confidence scores, with code.

Vision-language models (VLMs) are increasingly used for medical image interpretation, yet they frequently hallucinate, generating clinically plausible but factually incorrect findings that pose direct patient safety risks. We introduce HalluCXR, a benchmark evaluating six architecturally diverse VLMs across 856 stratified MIMIC-CXR chest radiographs and three query types, yielding 15,408 model evaluations. An eight-category hallucination taxonomy with clinical severity ratings and a two-layer detection pipeline are validated against 250 human annotations (auto-detection F1=0.959; LLM judge F1=0.907). We find that 61.9--82.3% of outputs contain hallucinations, with clinically dangerous errors in up to 80.2%. Three key patterns emerge: normal radiographs paradoxically attract the most severe hallucinations, common findings are systematically over-fabricated while rare findings go under-detected, and response length alone predicts hallucination risk (AUC up to 0.908). A six-model ensemble reduces fabrication by up to 84.8% at the cost of increased omission; a three-model subset retains comparable performance at half the cost. These results establish that hallucination auditing, verbosity-based risk monitoring, and ensemble-based safety layers are prerequisites for clinical deployment.