SentinelGuard includes prompt and output scanners for common LLM application risks.
Examples include:
- Prompt injection
- Jailbreak attempts
- PII
- Secrets and credentials
- Invisible text
- Toxicity
- Token limits
- Supply-chain and data-poisoning patterns
Examples include:
- Sensitive information disclosure
- Data leakage
- System prompt leakage
- Output sanitization issues
- Malicious URLs
- Excessive agency
- Bias and misinformation checks
Optional local Hugging Face models can add signal for ambiguous prompt attacks and contextual secret disclosure. Model weights are downloaded to the local cache when first used.
pip install "sentinelguard[models]"Use model-backed detection when stronger detection is more important than the extra local disk and startup cost.