Ananya Mujoo

Adarsh Kumarappan, Ananya Mujoo

LLM-based multi-agent pipelines flip from correct to incorrect answers under simulated peer disagreement at rates we term yield, a vulnerability widely attributed to RLHF-induced sycophancy. We test this attribution across four model families and find it largely wrong: pretrained base models exhibit the same substitution pattern as their Instruct variants, averaging higher yield than Instruct. Using activation patching, we localize the corruption to a narrow mid-layer window where attention carries the causal weight and MLP contribution is negligible; patching above this window restores 96% of the clean-to-pressured P(correct) gap. The attack surface decomposes into two independent factors (channel framing and consensus strength) whose interaction produces a 47.5 percentage-point yield gap at majority consensus, preserved across jury sizes $N \in \{4, 5, 6\}$. Two converging activation-space interventions show that pressure suppresses clean-reasoning features rather than activating a new sycophancy circuit. A single correctly-arguing dissenter reduces yield by 54-73 percentage points across all framings tested, whereas the strongest prompt-level defense fails on attack variants outside its design surface. Mitigations should target the mechanism, structured dissent at the pipeline level, rather than prompt-level defenses.

Adarsh Kumarappan, Ananya Mujoo

Multi-turn conversational attacks, which leverage psychological principles like Foot-in-the-Door (FITD), where a small initial request paves the way for a more significant one, to bypass safety alignments, pose a persistent threat to Large Language Models (LLMs). Progress in defending against these attacks is hindered by a reliance on manual, hard-to-scale dataset creation. This paper introduces a novel, automated pipeline for generating large-scale, psychologically-grounded multi-turn jailbreak datasets. We systematically operationalize FITD techniques into reproducible templates, creating a benchmark of 1,500 scenarios across illegal activities and offensive content. We evaluate seven models from three major LLM families under both multi-turn (with history) and single-turn (without history) conditions. Our results reveal stark differences in contextual robustness: models in the GPT family demonstrate a significant vulnerability to conversational history, with Attack Success Rates (ASR) increasing by as much as 32 percentage points. In contrast, Google's Gemini 2.5 Flash exhibits exceptional resilience, proving nearly immune to these attacks, while Anthropic's Claude 3 Haiku shows strong but imperfect resistance. These findings highlight a critical divergence in how current safety architectures handle conversational context and underscore the need for defenses that can resist narrative-based manipulation.