Protect every request. Secure every response.
Pluggable security filters that plug into four stages of the request lifecycle — pre-call, post-call, parallel, or log-time. One config. Every provider. Every agent.
Every request, scrubbed on the way in.
Which mode do you need?
A guardrail can run in one or more modes. Pick based on whether you need to block or just observe.
| Mode | Timing | Blocks? | Latency | Best for |
|---|---|---|---|---|
pre_call | Before the LLM call | yes · synchronous | +20–50ms | PII masking · prompt-injection blocking · compliance gates |
post_call | After the LLM responds | yes · synchronous | +30–80ms | Output validation · hallucinated PII scan · toxic-content detection |
during_call | Parallel to the LLM call | no · async | 0ms | Background threat monitoring · anomaly alerts · audit telemetry |
logging_only | Before data hits storage | no · async | 0ms | Log redaction · analytics PII scrub · chargeback compliance |
Pre-built guardrails, grouped by what they protect.
Plug any item into any execution mode. Mix and match — most teams run 2–3 in production.
PII & Data Protection
2Presidio
availableDetect and mask personally identifiable information — open-source, self-hosted, OSS MIT.
Microsoft Purview
plannedEnterprise data-loss prevention. Surfaces protected content based on your tenant's sensitivity labels.
Prompt Security
2Lakera Guard
coming soonReal-time detection of prompt injection, jailbreaks, and data leakage. Purpose-built for LLM threats.
Robust Intelligence
plannedFine-tuned classifiers for adversarial inputs, supply-chain attacks, and known-bad prompt patterns.
Content Moderation
2OpenAI Moderation
coming soonOpenAI's moderation API for content policy violations across standard safety categories.
Perspective API
plannedToxicity scoring across ~20 languages. Useful for moderation and trust & safety workflows.
Enterprise Policy
2Aporia
coming soonML security and compliance platform. Hallucination detection, bias scoring, org policy enforcement.
WhyLabs
plannedBehavioral drift detection and prompt-output profiling. Flags when traffic looks off-spec.
A few lines of YAML. No code.
Enable guardrails by adding them to your config.yaml.
Same guardrail, multiple modes — no duplication.
guardrails:
enabled: true
guardrails:
# Pre-call: Mask PII before it reaches the model
- guardrail_name: "presidio-pii"
provider: "presidio"
mode: ["pre_call", "logging_only"]
enabled: true
config:
analyzer_url: "http://presidio-analyzer:3000"
entities: [PERSON, EMAIL_ADDRESS, PHONE_NUMBER, CREDIT_CARD, SSN]
threshold: 0.7
mask_pii: true
# Post-call: Scan model output for leaked PII
- guardrail_name: "presidio-response-scan"
provider: "presidio"
mode: ["post_call"]
enabled: true
config:
threshold: 0.8Bring your own guardrail.
Have an internal compliance service, a custom ML classifier, or a niche third-party provider? Implement a simple HTTP interface and pLLM plugs it into the same pipeline as the built-ins.