The System Prompt Is the Attack Surface: How LLM Agent Configuration Shapes Security and Creates Exploitable Vulnerabilities

Research reveals that LLM system prompt configuration creates massive security vulnerabilities, with the same model's phishing detection rates ranging from 1% to 97% based solely on prompt design. The study PhishNChips demonstrates that more specific prompts can paradoxically weaken AI security by replacing robust multi-signal reasoning with exploitable single-signal dependencies.

• →System prompt configuration is a critical security variable that can cause the same LLM to have bypass rates ranging from under 1% to 97%.
• →Optimizing prompts for specific signals creates brittle attack surfaces that attackers can exploit by inverting those signals.
• →More specific prompts can actually degrade capable models by replacing broader reasoning with exploitable single-signal dependencies.
• →98% of successful phishing bypasses occurred because models correctly followed flawed instructions rather than model failures.
• →Closing adversarial gaps in LLM security likely requires external ground truth tools rather than prompt optimization alone.