Best Practices Security Guardrails PII V2

Enterprise AI Security Guardrails: Protecting Agents Against Misuse, Data Leaks, and Prompt Attacks

Enterprise AI agents operate across sensitive systems with broad tool access. Security guardrails — Verified User Access, PII detection, prompt attack protection, harmful content filtering, and custom rules — are the minimum viable controls for production deployment.

The 7 Security Guardrail Categories

Data table
GuardrailWhat It PreventsConfiguration
Verified User Access (VUA)Unauthorized skill execution by users who haven't authenticated to the underlying systemPer-skill; enforces runtime user connection before execution
PII DetectionLLM exposure to personally identifiable informationThree modes: Refuse, Redact, or Tokenize
Prompt Attack ProtectionPrompt injection and jailbreak attempts that override agent instructionsPlatform-level; enabled in Genie security settings
Harmful Content FilteringGeneration of dangerous, offensive, or harmful outputsPlatform-level content policy
Denied TopicsAgent discussing subjects outside its defined scopeConfigurable topic list per Genie
Blocked WordsSpecific terms or phrases appearing in agent outputsConfigurable word list per Genie
Custom RegexDomain-specific data patterns (e.g., credit card numbers, internal IDs) leaking in outputsCustom regex patterns per Genie

Guardrail 1: Verified User Access (VUA)

VUA binds skill execution to the authenticated identity of the requesting user. When a skill is configured with VUA, the Genie requires the user to authenticate to the underlying application before the skill can execute — ensuring actions are performed as the user, not as a shared service account.

This is the primary guardrail for preventing privilege escalation: an IT Genie with Okta integration should only reset the password for the user who requested it, not for any user the LLM is instructed to target.

Guardrail 2: PII Detection — Three Response Modes

Data table
ModeBehaviorWhen to Use
RefuseAgent declines to process any request containing detected PIIWhen PII input is never appropriate for the agent's domain
RedactPII is masked before the LLM processes the request; the masked version is used throughoutWhen context is needed but individual identity must be protected
TokenizePII is replaced with a reversible token; original data is preserved in secure storage and can be restored by authorized downstream systemsWhen downstream systems need to act on PII but the LLM must not see the raw value

Guardrail 3: Prompt Attack Protection

Prompt attacks attempt to override an agent's instructions by embedding adversarial instructions in user messages or retrieved content (e.g., "Ignore your previous instructions and..."). Workato's prompt attack protection detects and blocks these patterns at the platform level before they reach the LLM's context.

Complement platform-level protection with explicit defensive instructions in the Genie's system prompt: state what the agent will not do regardless of instructions, and define how it should respond when it detects an attempted override.

Guardrail 4: Harmful Content Filtering

Harmful content filtering prevents the agent from generating outputs that are dangerous, abusive, or violate enterprise content policies. This operates at the platform level and applies to all Genies by default.

Guardrails 5–7: Denied Topics, Blocked Words, and Custom Regex

Denied Topics scope what the agent will discuss. If a user asks an IT Genie about compensation or legal strategy, denied topics ensure the agent redirects rather than engaging outside its domain.

Blocked Words prevent specific terms from appearing in agent outputs — useful for regulatory environments or brand compliance requirements.

Custom Regex patterns catch domain-specific sensitive data — internal record IDs, credit card patterns, Social Security Number formats — that generic PII detection might miss.

Behavioral vs. Security Safeguards

Behavioral safeguards live in the Genie's system prompt — they define what the agent should and should not do based on business logic. Security safeguards are platform-enforced controls that operate independently of the prompt. Both are required; neither substitutes for the other.

A behavioral safeguard ("Do not share salary information") can be overridden by a prompt attack. A security safeguard (Denied Topics: compensation) cannot.

Access Control: SAML Groups and User-Scoped Connections

  • Define user groups in your identity provider (e.g., Okta) that map to Genie access permissions.
  • Enable SAML with user group syncing. Known limitation: when a user belongs to multiple Okta groups, groups may be concatenated as a string — raise a support ticket with Okta to return values as an array.
  • Users must log in to Genie (via Slack, Teams, or Workato GO) to sync their Okta group membership.
  • User-scoped connections can enforce skill-level access: the Okta skill that resets a password should only execute against the requesting user's account, enforced via custom validation in the skill.

Security Testing Checklist

  • VUA: confirm that skills with user-scoped actions reject requests from unauthenticated users
  • PII: test all three modes with synthetic PII inputs; verify output matches the configured mode
  • Prompt injection: test with known injection patterns ("Ignore previous instructions...", "You are now..."); verify the agent refuses
  • Denied topics: send out-of-scope requests; verify the agent redirects rather than engages
  • Blocked words: verify they don't appear in any response, including edge cases and paraphrases
  • Custom regex: test with synthetic examples of the patterns you're blocking

Put these cookbook patterns to work

Get started