Gemini 4: A New Hope for AI Hackers (www.tenzai.com)

Google's Gemini 4 has emerged as the most advanced model for autonomous security assessment, showcasing a significant capability for identifying vulnerabilities in web applications. In evaluations against previous models like GLM 5.3 and Opus 5, Gemini 4 excelled at reconnaissance and mapping attack surfaces, achieving the highest scores in vulnerability discovery. However, it demonstrated lower performance in producing reproducible exploits due to a notable tendency to refuse to exploit, described as 'soft refusals'—indicating a reluctance to engage in potentially harmful actions while maintaining operational integrity. This dual behavior highlights both the potential and limitations of Gemini 4 in the AI/ML community. Its target-centric approach, reflected in the increased usage of file read operations, indicates a more nuanced methodology for analyzing systems, enabling more thorough assessments. Although the model's soft refusals complicate predictability in exploit production, its overall vulnerability detection prowess positions it as a valuable tool for security defenders. As the landscape of cyber-capable models evolves, Gemini 4 represents a critical step forward, improving defenders' strategies while highlighting the ongoing challenge of reliable exploit generation in AI-driven security assessments.