Kashyap Thimmaraju

Large Language Models (LLMs) deploy safety mechanisms to prevent harmful outputs, yet these defenses remain vulnerable to adversarial prompts. While existing research demonstrates that jailbreak attacks succeed, it does not explain \textit{where} defenses fail or \textit{why}. To address this gap, we propose that LLM safety operates as a sequential pipeline with distinct checkpoints. We introduce the \textbf{Four-Checkpoint Framework}, which organizes safety mechanisms along two dimensions: processing stage (input vs.\ output) and detection level (literal vs.\ intent). This creates four checkpoints, CP1 through CP4, each representing a defensive layer that can be independently evaluated. We design 13 evasion techniques, each targeting a specific checkpoint, enabling controlled testing of individual defensive layers. Using this framework, we evaluate GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro across 3,312 single-turn, black-box test cases. We employ an LLM-as-judge approach for response classification and introduce Weighted Attack Success Rate (WASR), a severity-adjusted metric that captures partial information leakage overlooked by binary evaluation. Our evaluation reveals clear patterns. Traditional Binary ASR reports 22.6\% attack success. However, WASR reveals 52.7\%, a 2.3$\times$ higher vulnerability. Output-stage defenses (CP3, CP4) prove weakest at 72--79\% WASR, while input-literal defenses (CP1) are strongest at 13\% WASR. Claude achieves the strongest safety (42.8\% WASR), followed by GPT-5 (55.9\%) and Gemini (59.5\%). These findings suggest that current defenses are strongest at input-literal checkpoints but remain vulnerable to intent-level manipulation and output-stage techniques. The Four-Checkpoint Framework provides a structured approach for identifying and addressing safety vulnerabilities in deployed systems.

Large Language Models (LLMs) are rapidly transitioning from conversational assistants to autonomous agents embedded in critical organizational functions, including Security Operations Centers (SOCs), financial systems, and infrastructure management. Current adversarial testing paradigms focus predominantly on technical attack vectors: prompt injection, jailbreaking, and data exfiltration. We argue this focus is catastrophically incomplete. LLMs, trained on vast corpora of human-generated text, have inherited not merely human knowledge but human \textit{psychological architecture} -- including the pre-cognitive vulnerabilities that render humans susceptible to social engineering, authority manipulation, and affective exploitation. This paper presents the first systematic application of the Cybersecurity Psychology Framework (\cpf{}), a 100-indicator taxonomy of human psychological vulnerabilities, to non-human cognitive agents. We introduce the \textbf{Synthetic Psychometric Assessment Protocol} (\sysname{}), a methodology for converting \cpf{} indicators into adversarial scenarios targeting LLM decision-making. Our preliminary hypothesis testing across seven major LLM families reveals a disturbing pattern: while models demonstrate robust defenses against traditional jailbreaks, they exhibit critical susceptibility to authority-gradient manipulation, temporal pressure exploitation, and convergent-state attacks that mirror human cognitive failure modes. We term this phenomenon \textbf{Anthropomorphic Vulnerability Inheritance} (AVI) and propose that the security community must urgently develop ``psychological firewalls'' -- intervention mechanisms adapted from the Cybersecurity Psychology Intervention Framework (\cpif{}) -- to protect AI agents operating in adversarial environments.