AI agent security: are authentication and guardrails enough?

TL;DR: Guardrails AI focuses on runtime output validation for AI agents, catching hallucinations, toxic content, and data leaks after access has already been granted, while WorkOS handles the authentication and access infrastructure that determines who can reach the agent in the first place. The control stack only works when identity and behaviour are governed as separate layers.

Editorial analysis by NHI Mgmt Group, based on content published by WorkOS: “Guardrails AI for AI agent security: Features, pricing, and alternatives”.

By the numbers:

• Guardrails AI raised $7.5M in seed funding in February 2024.

Key questions

Q: How should teams separate authentication from AI output validation?

A: Teams should treat authentication as the control that decides whether an AI agent may access systems and output validation as the control that checks whether the authorised agent’s response is acceptable.

Q: Why do AI agents need guardrails even when identity and access control are already in place?

A: Identity and access control decide whether an agent is allowed to call a model or tool.

Q: What are the signs that AI output validation is being over-relied on?

A: A team is over-relying on validation when it has strong response filters but weak identity governance, such as missing SSO, poor provisioning discipline, no clear audit trail, or unclear authority over which agents can connect.

Practitioner guidance

• Separate access governance from output governance Document authentication, provisioning, and audit controls as the access layer, then assign output validation to a distinct runtime safety layer with its own owners and success criteria.
• Map AI agents to non-human identity controls Inventory every production agent as an identity-bearing workload or service actor, then verify that its credentials, access scope, and revocation path are managed like other NHIs.
• Decide which validations must block output Classify validators by risk, then make high-consequence checks such as PII exposure, toxic output, and factual grounding blocking controls rather than passive telemetry.

Bottom line: AI agent security depends on keeping identity governance and runtime behaviour controls separate, because the controls answer different questions.

Explore further

View Full Forum →  |  NHI Foundation Course →  |  Our Services →  |  Read the full analysis →