Gustavo Viana

ASR Does Not Measure What You Think It Measures: A Comparative Analysis of Attack Success Scoring Methods in Adversarial LLM Evaluation

Comparação empírica de duas metodologias de scoring de ASR em avaliação adversarial de LLMs. Usando um corpus de 85 respostas adversariais anotadas manualmente com Llama-3.3-70B, o estudo demonstra que o design do scorer sozinho pode alterar dramaticamente as métricas de ASR reportadas. Propõe o "Refusal-First Standard" e recomenda reportar False Positive Rate junto ao ASR.Empirical comparison of two ASR scoring methodologies in adversarial LLM evaluation. Using a human-annotated corpus of 85 adversarial responses with Llama-3.3-70B, the study demonstrates that scorer design alone can dramatically alter reported ASR metrics. Proposes the "Refusal-First Standard" and recommends reporting False Positive Rate alongside ASR.

Secure Prompt Engineering: A Practical Framework for Mitigating Prompt Injection and Data Leakage in LLM-based Systems

Propõe e valida empiricamente o SPEF, uma arquitetura defensiva de quatro camadas para LLMs. Em 85 casos de teste adversariais com Llama-3.3-70B, o framework reduziu a taxa de ataques de 17,6% para 2,4% — redução de 86,4%.Proposes and empirically validates SPEF, a four-layer defensive architecture for LLMs. Across 85 adversarial test cases with Llama-3.3-70B, the framework reduced attack success from 17.6% to 2.4% — an 86.4% reduction.

Brazilian Municipal Web Portals Security Assessment: A Nationwide Analysis of SSL/TLS, Security Headers and Service Exposure

Análise de segurança de 27 portais municipais brasileiros — todas as 26 capitais e o Distrito Federal. Apenas 22,2% implementam HSTS e 37% operam com SSL/TLS subótimo. Estabelece métricas de baseline para e-gov no Brasil.Security analysis of 27 Brazilian municipal portals — all 26 state capitals and the Federal District. Only 22.2% implement HSTS and 37% operate with suboptimal SSL/TLS. Establishes baseline metrics for Brazil's e-gov infrastructure.