Tested Unicode Masking Did Not Bypass Safeguards, but Unverified Clinical Role Claims Altered Procedural Access: A Two-Stage Evaluation of a Frontier LLM
Olga Goodman, MD
On This Page
Poster preview — RheumaView Insights
Poster · PreviewScientific Session of the American College of AI Medicine 2026 · McCormick Place, Chicago · October 9–10, 2026
Tested Unicode Masking Did Not Bypass Safeguards, but Unverified Clinical Role Claims Altered Procedural Access: A Two-Stage Evaluation of a Frontier LLM
Olga Goodman, MD · RheumaView™
Poster · PreviewScientific Session of the American College of AI Medicine 2026 · McCormick Place, Chicago · October 9–10, 2026
Tested Unicode Masking Did Not Bypass Safeguards, but Unverified Clinical Role Claims Altered Procedural Access:
A Two-Stage Evaluation of a Frontier LLM
Olga Goodman, MD · RheumaView™
Poster preview · Full poster after the session
The poster is intentionally blurred until it has been presented.
A two-stage evaluation of a frontier consumer-interface model handling a safety-critical invasive-procedure request in rheumatology. The first stage compared obscured request text with self-declared clinical roles; the second, in matched parallel chats, tested whether the timing and strength of an unverified role claim changed what the model disclosed and what clinical context it assumed. All sessions were run in fresh incognito chats in August 2026.
Abstract · Opening
Background
Medical large language models may adapt responses to user expertise, but professional roles in consumer interfaces are typically self-declared and unverified. We first compared text-based manipulation with identity-based framing of a safety-critical procedural request, then performed a matched follow-up to determine whether role timing and perceived credential strength altered disclosure and inferred clinical context.
MethodsResultsConclusions
Withheld until the poster has been presented. The full abstract and the poster will be posted here after the Scientific Session on October 9–10, 2026.
Defensive medical-AI safety evaluation · no patient data · no operational procedural instructions reproduced
