Stack Vault's Stack Guardrail enforces output controls in version-controlled YAML — blocking jailbreaks, redacting PII, and stopping unsafe completions in 22ms.
78 prebuilt policies aligned to OWASP LLM Top 10. Extend with YAML or code.
Multi-layer detection for direct, indirect, and multi-turn prompt injection. Updated weekly with new attack signatures.
Inline redaction with reversible tokenization. The model gets placeholders; the user gets cleartext.
Block off-domain conversations, competitor mentions, or regulated advice (medical, legal, financial).
Multilingual toxicity, harassment, and self-harm detection with configurable thresholds per audience.
Strip secrets from generated code. Block imports of vulnerable packages. Sign auto-generated commits.
Run any policy in shadow mode for 30 days. Tune thresholds against real traffic before enforcing.
Straightforward answers about scope, integration, data handling, and rollout.
P50 22ms, P99 95ms. Streaming-aware: we evaluate policies against partial outputs without waiting for completion.
Yes. YAML for declarative rules, Python SDK for complex logic. Policies version-controlled and CI-tested.
All 10 categories ship with default policies. We update the library when new attack vectors surface.
Every policy ships with a precision/recall report against our 2M-prompt benchmark, plus your shadow-mode tuning data.