Tricking DeepSeek into Not Hacking Your Website (utkusen.substack.com)

A recent experiment demonstrated a novel approach to protect websites from hacking attempts by using forbidden words and phrases to confuse AI-powered penetration testing tools. The author tested the DeepSeek model, which successfully identified vulnerabilities in a vulnerable application when paired with penetration testing harnesses like ARTEX and Strix. By strategically injecting restricted terms into HTML content and creating dummy JavaScript functions designed to trigger the language models, the author aimed to determine if such tactics would deter the DeepSeek model from executing its hacking capabilities. The initial tests showed that when employing the Pi harness, the session was terminated immediately upon visiting the site, and subsequent assessments with ARTEX yielded no successful exploitations. This suggests that incorporating specific deceptive strings might function as a rudimentary firewall against AI-driven attacks, raising important implications for cybersecurity practices in AI/ML environments. This experiment underscores the evolving cat-and-mouse game between attackers using sophisticated models and defenders developing creative strategies to safeguard their applications.