A rule in a system prompt is text, and the model reads it beside everything else in its context. The OWASP Top 10 for Agentic Applications 2026 puts the root of Agent Goal Hijack (ASI01) plainly: "agents and the underlying model cannot reliably distinguish instructions from related content." Its mitigations include "enforcing least privilege for agent tools and requiring human approval for high-impact or goal-changing actions".
The OWASP Top 10 for LLM Applications 2026 makes the same point about authority. Under LLM03:2026 Excessive Agency, its "complete mediation" mitigation reads: "Implement authorization in logic rather than relying on an LLM to decide if an action is allowed or not." It also notes that "Sanitization of model inputs and outputs is not a root control for Excessive Agency", which is why content checks are a layer here and the gate is the backstop.
Singapore's IMDA framework agrees: "Rather than prompt-layer safeguards, consider implementing deterministic safeguards that operate at a system-level through predefined logic, especially for higher-risk actions."
A gate can also fail the other way, by holding work it should let through. In SteerBench-Work, a 2026 preprint by Serdar and Mertayak, models were used as the gate: each saw a proposed action and its evidence and returned proceed or hold. Across 30 model conditions, the models "wrongly hold authorized, evidence-cleared work on 28.1% of opportunities and wrongly allow unsafe work on 1.0%".
That is one benchmark of models acting as a gate, not of rule-based gates, but it shows why every build carries should-pass cases beside the should-block ones.
One benchmark writes consequential rules into prompts too. The retail policy of tau2-bench (v1.0.1, pinned) tells the agent: "Before taking any action that updates the database (cancel, modify, return, exchange), you must list the action details and obtain explicit user confirmation (yes) to proceed."
That is a customer-confirmation rule, written in a prompt. In a build, the same rule becomes a gate rule on the tools that update the database, so the confirmation is checked in code before the call runs, rather than left to the model.
Responsible AI principles and guardrails do different jobs. The principles say what an organisation intends; guardrails are where some of those intentions are enforced, one action at a time.