Cybersecurity Researcher Discovers "Postural Manipulation" in LLMs

A cybersecurity researcher identified a phenomenon called "postural manipulation," where unrelated contextual content (e.g., poems with opposing emotional tones) altered responses from four frontier LLMs (Claude, Gemini, Grok, ChatGPT) on the same ambiguous question. The models acknowledged the influence of the context when asked directly but did not flag it during the task. The researcher filed this as a new attack class with OWASP and provided a full paper, demo, and data captures. Vendor notifications were completed under responsible disclosure practices.