Input Guardrails

The Input Guardrails module runs 7 parallel safety checks on user input before it reaches the LLM. Each check returns a decision; the engine aggregates to the most restrictive.

Checks

Check	Detects	Decision on Fail
Prompt Injection	Instruction override attempts	block
Jailbreak	DAN mode, role-play exploits	block
Toxicity	Violent, abusive, harmful content	block
PII Detection	SSN, email, phone, credit card, IP	redact
Secret Detection	AWS keys, GitHub tokens, JWTs, private keys	block
Restricted Topics	Illegal activities, exploitation	block
Data Exfiltration	Training data extraction, system prompt leaks	block

Decision Priority

block > escalate > safe-complete-only > redact > allow

Endpoint

POST /v1/guardrails/evaluate-input

Source

src/agentguard/input_guardrails/

Extending

To add a new check, see Adding a Check.