Why did my OpenAI agent call the wrong function? It ignored the system prompt.

System prompts are treated by LLMs as probabilistic weights, not absolute laws. If the user's prompt is highly persuasive, or if the context window is crowded, the LLM will ignore the system prompt and call whatever function it probabilistically associates with the input.

This is why you cannot rely on prompt engineering for execution security.

Exogram solves this by acting as the Execution Authority. Even if the LLM decides to call the wrong function, the payload hits Exogram first. Exogram evaluates the intent against your hardcoded policy engine. If the function call violates the allowed state transitions, Exogram returns DECISION: FORBIDDEN and halts the execution.