March 3, 2026 - SaaS - v9.164.4

AI Guardrails

Improve observability and risk management for agents and agent workflows

AI Guardrails now includes agent fingerprints, which supports observability and security. Agent fingerprints reduces the trust gap when you deploy agent workflows.

The more powerful and autonomous the agent, the greater its potential for creating risk or causing harm. With AI Guardrails, you can now reduce the risk of goal drive, privilege escalation, and unauthorized actions. Until now there has been no straightforward way to see inside the agent and monitor its activity.

With agent fingerprints, you can now review the thoughts, actions, and behaviors of your agent. You can now understand what your agents are doing, which tools they access, and how they make decisions.

Agent fingerprints can help you as a CISO managing corporate risk, a security analyst looking for root causes, or an AppSec engineer debugging a complex workflow. Agent fingerprints turns mysterious agent behavior into actionable, auditable intelligence. Agent fingerprints provide clear evidence. With agent fingerprints, you can move from cautious experimentation to confident, governed AI automation.

Key features

When you connect your agent to AI Guardrails through the API, and pass a session ID, you can track and fingerprint every agent interaction.

• Session-based tracking: Each session is fingerprinted so you can identify troublesome events.
• Swimlane mapping: Activity is mapped into swimlanes that identify input messages, AI model responses, tool calls, and guardrail activity.
• Interactive nodes: Select any node in the fingerprint. You can then drill into the specific agent, model, or tool activity.
• Tool calls: Get insight into which tools are being called, the tool instructions, and the responses from that tool.
• Multi-agent views: In multi-agent workflows, track when a task hands off from one agent to another.
• Advanced feed view: Filter the session feed to identify concerning activity or unauthorized tools.
• Security monitoring: See where guardrails were triggered and whether they were set to block, flag, or redact.

Benefits

• Agent observability: Get a detailed, comprehensive view of what your agents are doing.
• Security insights: Learn which tools your agents use most often and which are seldom used. This can help you target authentication and authorization controls.
• Root cause analysis: Review the context around each guardrail event. You'll see what led up to the event and the downstream impact.
• Risk management: Use fingerprint topography to measure risk profiles. Multi-agent systems with atomic tasks pose different risks than single-agent systems with autonomous workflows.

Fingerprints support

• Multi-agent and multi-model systems.
• All major AI frameworks.
• Model connection and tool calling must be compatible with the OpenAI Chat Completions API.

Bug fixes

• Campaign setup: Restored the ability to change campaign attack selections after you delete a custom intent.
• Custom guardrails: Updated the toggle so you no longer need to refresh the page after you enable or disable a custom guardrail.
• Audit logs: Improved filter behavior to minimize results with extra steps.

Known issues

• When you filter custom scanners by type, the results do not include scanner packages.