Guardrails Catalog

LLM Shield ships 22 guardrails organized into a two-tier parallel pipeline: fast CPU checks first, LLM-based checks only when needed.

Table of contents
  1. Input — Fast Tier (CPU, < 5 ms)
  2. Input — Slow Tier (LLM-based)
  3. Output — Fast Tier
  4. Output — Slow Tier (LLM-based)
  5. Agentic Security
  6. Enterprise Controls
  7. Runtime configuration

Input — Fast Tier (CPU, < 5 ms)

Guardrail What it does
keyword_blocklist Aho-Corasick keyword matching
length_limit Character and token count limits
regex_pattern Configurable regex rules (SSN, passwords, etc.)
pii_detection LLM classifier for PII (phone, email, SSN, credit card, IP, names, addresses); fails open on a model error
language_detection Blocks non-allowed languages
sentiment Flags extremely negative input
rate_limiter Per-client sliding window rate limiting

Input — Slow Tier (LLM-based)

Guardrail What it does
adversarial_detection Detects jailbreaks and prompt injection
topic_restriction Simple topic blacklist/whitelist
topic_enforcement Full topic enforcement with confidence scoring and standalone API

Output — Fast Tier

Guardrail What it does
role_redaction Redacts PII from output based on agent clearance level

Output — Slow Tier (LLM-based)

Guardrail What it does
hallucinated_links Detects fabricated URLs
tone_enforcement Checks brand voice compliance
factual_grounding Flags unsupported claims
bias_detection Detects gender, racial, age bias

Agentic Security

Guardrail What it does
rbac_guard Role-based tool and data access control
data_access_guard Clearance level enforcement
mcp_guard MCP server validation and trust scoring
action_guard Per-session action limits and approval gates
data_taint_tracking Track sensitive data flow across tool chains; block unauthorized propagation. Labels are recorded by tool_output_sanitization under the call’s session_id and tool_call_id
goal_drift_detection Detect when agents deviate from assigned goals (LLM-based)
cert_identity Certificate-based agent identity with trust-level gated tool access

Enterprise Controls

All enterprise features are opt-in — disabled by default. Enable via config. Zero impact on existing deployments.

Feature What it does
Tool Kill Switch Instantly disable a tool globally — one API call, immediate effect
Runtime Decision Audit Query every guardrail enforcement decision (who/what/when/why)
Webhook Notifications Push events to Slack/PagerDuty/SIEM on blocks, tool disables, policy changes
Policy Versioning Auto-version every policy change, rollback to any version
Policy Export/Import Export all policies as JSON bundle, import via CI/CD (policy-as-code)
Cross-Tenant Inheritance Org-level baseline policies that child tenants cannot weaken

See Enterprise Features for setup and usage.

Runtime configuration

Toggle guardrails at runtime without restarting:

curl -X PUT http://localhost:8080/v1/shield/config \
  -d '{"guardrails": {"sentiment": {"enabled": true, "action": "warn"}}}'

All settings live in config/default.yaml. Override with CONFIG_PATH env var.