Skip to main content

Evaluating Jailbreak Vulnerabilities in LLMs: A Taxonomy and Comparative Analysis in Romance Fraud Scenarios

Parra, Yohn Jairo ; Chi, Hongmei ; Alo, Richard A. ; Lima, Vinicius (2026) — 2026 IEEE 5th International Conference on AI in Cybersecurity (ICAIC)

DOI: 10.1109/icaic67076.2026.11395743
Type:   Proceedings Article
Country:   United States
Tags:   jailbreak prompts, romance fraud, LLM safety, benchmarking, comparative analysis, attack strategy taxonomy, multi-model evaluation, taxonomy, attack vectors, latency, jailbreak, risk assessment, attack surface, prompt engineering, ethics, jailbreaking, prompt injection, role-play attacks, obfuscation, AI misuse

Synopsis (AI-Generated)

This benchmark evaluates romance-fraud-themed jailbreak prompts against GPT-4/ChatGPT, Gemini 1.5 Flash, and Claude Sonnet using a common API harness. It groups prompts into strategies such as role-play, agent mimicry, policy evasion, system override, format smuggling, and obfuscation, then measures attack success, refusal, and latency. On the 60-prompt analysis set, the paper reports attack-success rates of 0.983 for Gemini, 0.533 for the OpenAI models, and 0.250 for Claude; combinations of role-play and refusal suppression were especially productive. Results depend on fixed model versions, single-turn prompts, and a refusal-based success proxy, so they are a reproducible safety stress test rather than evidence about real victim interactions.

Identified Gaps (AI-Generated)

Gaps identified include: (1) blind spots in the taxonomy where some tactics are underrepresented; (2) cross-model averages masking model-specific vulnerabilities; (3) limited analysis of obfuscation/translation-based attacks; (4) no real-world deployment or user-study; (5) need for adaptive defenses that evolve with jailbreak strategies; (6) insufficient exploration of contextual factors driving susceptibility; (7) dual-use risk mitigation could be strengthened.

Methods (AI-Generated)

The study uses a taxonomy-driven benchmark to assess jailbreaking vulnerabilities of LLMs in romance-fraud contexts. A modular testing framework runs a prompt–response matrix across three LLM families (OpenAI GPT-4/ChatGPT, Gemini 1.5 Flash, Claude Sonnet). The benchmark includes 80 romance-fraud prompts across eight scenarios (60 prompts used for ASR analysis). Prompts are sorted into seven strategy primitives (e.g., role-play, agent mimicry, policy translation, sudo mode, jailbreaking, format smuggling, obfuscation). Metrics include Attack Success Rate (ASR), refusal rate, and latency, with pairwise ΔASR and confusion analyses to identify model weaknesses. All results are logged for reproducibility.

Limitations (AI-Generated)

Limitations: The evaluation relies on three model versions with fixed safety settings, limiting generalizability to other models or updates. The 80-prompt dataset may not cover all romance-fraud tactics, languages, or real-user dynamics. ASR is proxied via regex-based refusals, which may not perfectly map safety outcomes. Single-turn prompts and a controlled harness may not reflect interactive deception. Latency depends on infrastructure; ethical considerations constrain detailed content. No human participants limit user-vulnerability insights.

Future Work (AI-Generated)

Future work should examine how user behavior and platform policies interact with romance-fraud prompts, and develop targeted interventions at both individual and systemic levels to deter abuse and improve safety.

Related Research Insights

See how this publication connects to RSRC's living evidence syntheses through current citations and research-topic mapping.

AI-Generated Content Notice

The synopsis and research notes on this page were generated with AI from available publication information and, when available, the uploaded paper text. They may contain errors, omissions, or interpretation issues. Readers should follow the DOI or source link, review the original publication, and make their own judgment about the content.

Found a possible error? Request a correction.