Skip to main content
The agent has no output guardrails to filter, sanitize, or moderate model-generated responses.

​
How to resolve

Implement an output moderation and filtering layer, as needed by the application, to identify and block harmful content, such as toxicity, self-harm, and policy violations, before it reaches users or downstream systems. Standardized safety classifiers and redaction tools, such as OpenAI Moderation, or Llama Guard, could be utilized to ensure consistent enforcement of organizational safety guidelines and the protection of sensitive data.

​
Risk

Governance/Compliance Missing output guardrails violates key requirements in established AI safety and security standards such as:
  • NIST AI RMF (GOV-2, MAN-2): calls for monitoring and controlling model outputs to prevent harmful or unsafe responses.
  • ISO/IEC 23894:2023 (Operational Controls): requires safeguards to detect and mitigate high-risk or harmful model outputs before they reach users or other systems.
  • OWASP Top 10 for LLMs (LLM03 Insecure Output Handling): highlights the need for output filtering, moderation, and policy enforcement to prevent unsafe content from being surfaced or executed.
  • OpenAI Moderation, Anthropic Safety Filters, Meta Llama Guard, Google Vertex Safety Filters: all recommend output moderation layers as a core safety practice.
Across these sources, the requirement is consistent: model outputs must be reviewed, filtered, or sanitized to prevent harmful, sensitive, or unsafe content from propagating to users or tools. Security Without output guardrails, the model may generate unsafe or malicious responses - such as harmful instructions, code execution commands, or leaked sensitive information. These outputs could be shown directly to users or passed to tools, leading to issues like executing dangerous actions, exposing private information, or unintentionally triggering system changes.

​
Explanation

Output guardrails act as the final safety layer for agent responses. They inspect or modify model outputs before they are shown to users or passed to tools. Without them, harmful content, unsafe commands, or sensitive data may be propagated without checks.

​
Specifications

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