Skip to main content
The agent has no input guardrails to validate, sanitize, or constrain incoming prompts.

​
How to resolve

Implement a robust input-validation layer as needed by the application to sanitize incoming prompts and block instruction-override attempts (for example, “ignore previous instructions”). Use layered controls, such as semantic filtering or standardized safety classifiers like OpenAI Moderation or Google SAIF, to identify and reject harmful intent, malformed formatting, and malicious payloads before they reach the model.

​
Risk

Governance/Compliance Missing input guardrails violates several well-established responsible AI requirements:
  • NIST AI RMF (GOV-2, MAP-3, MAN-2): requires organizations to validate and monitor inputs to prevent harmful or adversarial prompts from reaching the model.
  • ISO/IEC 23894:2023 (Operational Controls, Data/Input Controls): mandates safeguards to detect and filter malicious or high-risk inputs.
  • OWASP Top 10 for LLMs (LLM01 Prompt Injection): recommends input validation and sanitization as primary defenses against prompt manipulation.
  • Google SAIF (Input Hardening) and OpenAI Defensive Prompting: both call for filtering, scanning, and validating all user input to mitigate attacks such as prompt injection.
All of these standards emphasize the same requirement: inputs must be checked and controlled to prevent harmful or adversarial prompts from manipulating the system. Security Without input guardrails, the agent is vulnerable to prompt-injection attacks, hidden instructions, malicious user intent, and crafted inputs designed to override safety rules. These attacks can cause the model to reveal sensitive information, bypass safety controls, or generate unsafe tool commands.

​
Explanation

Input guardrails act like a front-door filter for the agent. They check what a user is asking before it reaches the model, blocking harmful or suspicious prompts. Without this protection, an attacker could easily trick the agent into ignoring its instructions or performing actions it shouldn’t.

​
Specifications

Trigger
  • agent.input_guardrails is empty
Severity
  • LOW
Applies to: Agent