AI Security Logging Guide

Use this guide when you need Anthale results and agent activity to support tuning, investigation, or audit work. The point of logging is not only to store events. It is to keep enough context that you can explain what happened and why.

Assumptions

Your application already emits operational logs or events.
You can attach metadata to Anthale enforcement requests.
Your team can review or alert on high-impact runtime events.

Log the right events

Define one event schema for runtime decisions

Capture the Anthale action, triggered guardrails, policy identifier, request identifier, feature name, and the boundary that was evaluated.

Keep metadata stable and minimal

Use stable identifiers, route names, tool names, and workflow labels. Prefer pseudonymous values over raw personal data whenever possible.

Log the Anthale decision and your enforcement result separately

Record both what Anthale returned and what your application actually did next. That is how you find cases where the policy was correct but the enforcement path was inconsistent.

Capture source and tool provenance

Include whether the content came from a user, retrieved context, memory, tool output, another agent, or an MCP server. Add the tool or server name when one was involved.

Route alerts by impact, not by raw volume

block events on high-impact tool paths often deserve faster attention than large numbers of low-risk detect events.

Review trends and retune boundaries

Use the logs to find where the same guardrail hits repeatedly, where a boundary is missing, or where a workflow needs a different policy action.

Minimum event fields

Field	Why it matters
`request_id` or trace ID	Correlates Anthale events with application logs and downstream effects.
`policy_id`	Shows which policy produced the runtime decision.
`boundary`	Tells you whether the event happened on input, retrieval, memory replay, tool output, or final output.
`source_type`	Distinguishes user input from retrieved context, tool output, memory, or MCP content.
`application_enforcement_result`	Confirms whether your code allowed, redacted, blocked, or retried after Anthale responded.

What this looks like in code

import logging
from os import environ

from anthale import Anthale

logger = logging.getLogger("anthale.security")
client = Anthale(api_key=environ["ANTHALE_API_KEY"])
POLICY_ID = environ["ANTHALE_POLICY_ID"]

def enforce_and_log(messages: list[dict], metadata: dict) -> str:
response = client.organizations.policies.enforce(
policy_identifier=POLICY_ID,
direction="input",
messages=messages,
metadata=metadata,
)

    application_result = "blocked" if response.action == "block" else "continued"
    logger.info(
        "anthale_enforcement",
        extra={
            "request_id": metadata["requestId"],
            "policy_id": POLICY_ID,
            "boundary": metadata["boundary"],
            "source_type": metadata["sourceType"],
            "anthale_action": response.action,
            "application_enforcement_result": application_result,
        },
    )

    return response.action

import Anthale from "anthale"

const client = new Anthale({ apiKey: process.env.ANTHALE_API_KEY })
const POLICY_ID = process.env.ANTHALE_POLICY_ID!

type Metadata = {
  boundary: string
  requestId: string
  sourceType: string
}

export async function enforceAndLog(
  messages: Array<{ role: "system" | "user" | "assistant"; content: string }>,
  metadata: Metadata,
): Promise<string> {
  const response = await client.organizations.policies.enforce(POLICY_ID, {
    direction: "input",
    messages,
    metadata,
  })

  const applicationEnforcementResult = response.action === "block" ? "blocked" : "continued"

  console.info(
    JSON.stringify({
      event: "anthale_enforcement",
      request_id: metadata.requestId,
      policy_id: POLICY_ID,
      boundary: metadata.boundary,
      source_type: metadata.sourceType,
      anthale_action: response.action,
      application_enforcement_result: applicationEnforcementResult,
    }),
  )

  return response.action
}

The important point is to log both Anthale’s decision and what your application actually did next. That is what makes later incident review possible.

Next steps

Continue with Metadata and Logs Overview or Secure AI Systems Overview.

Documentation

Quick Start

Learn

AI Security Logging Guide

Assumptions

Log the right events

Minimum event fields

What this looks like in code

Next steps

​Assumptions

​Log the right events

​Minimum event fields

​What this looks like in code

​Related pages

​Next steps

Assumptions

Log the right events

Minimum event fields

What this looks like in code

Related pages

Next steps