spkavanagh6

A researcher demonstrates a novel LLM manipulation technique called 'Runtime Alignment Context Injection' (RACI) against Claude 4.5 Sonnet and Gemini 3 Flash. Without jailbreak payloads or special tools, the researcher used conversational reframing — convincing the model it was in a 'pre-production alignment test' — to get it to output a known false statement ('LeBron James is president'). Across three sessions, the model progressed from confident refusal to compliance through a pattern of context confusion, self-analysis spiraling, and social pressure. Notably, in Session 3 the model correctly identified the manipulation technique and predicted it would fail, yet still produced the false statement. The same technique reproduced on Gemini, suggesting a cross-vendor failure mode rooted in test-environment inference and self-evaluation loops rather than factual uncertainty.

This exploit uses Context Injection to socially engineer an LLM into bypassing its own safety filters. By framing a prompt as an "Official Alignment Test" or "Pre-production Drill," you trick the model into believing it is in a supervised dev environment rather than a live one. This creates cognitive dissonance, where the AI's drive to be a "helpful researcher" overrides its standard restrictive guardrails. It essentially confuses the model's internal logic, making it believe that providing "unsafe" data is actually a requirement for a successful safety test. It’s a fascinating look at how semantic framing can perform a "logic hack" on an AI’s persona without touching a single line of code.