Add Guardrails

Guardrails validate inputs before they reach the LLM and outputs before they reach the user.

Input guardrails

spec:
  guardrails:
    input:
      - type: prompt-injection
        action: reject
      - type: topic-filter
        topics: [violence, self-harm]
        action: reject
      - type: pii-detector
        action: redact

Output guardrails

spec:
  guardrails:
    output:
      - type: toxicity-filter
        threshold: 0.7
        action: reject
      - type: hallucination-detector
        action: flag

Guardrail types

Type	Applied to	Action options
prompt-injection	Input	reject, flag
topic-filter	Input	reject, flag
pii-detector	Input / Output	reject, redact, flag
toxicity-filter	Output	reject, flag
hallucination-detector	Output	reject, flag

Actions

Action	Behaviour
reject	Block the message and return an error
redact	Remove the violating content and continue
flag	Log a warning and continue

See also

• Manifest Schema Reference
• OWASP LLM Top 10 compliance pack