Say Please (if only as a reminder)

In a recent article on Dev.to, developer Bill Tarbox explores the limitations of LLM prompt engineering for security. He argues that system prompts act as polite requests rather than deterministic controls, comparing them to a 'Please Do Not Touch' sign in a museum. Because LLMs are probabilistic, they can be bypassed by persuasive users. Tarbox suggests a three-layered security approach: the 'sign' (system prompts for behavior), the 'guard' (external guardrails and classifiers), and the 'glass' (programmatic determinism). He emphasizes that developers should never rely on prompts for critical security tasks. Instead, developers must implement hard constraints, such as least-privilege IAM roles and input validation, to ensure safety. Ultimately, the author advises that while prompts are useful for shaping tone and general behavior, developers must rely on code to enforce security boundaries, treating prompts as requests and code as the final, unbreakable authority.