A Security Field Guide to AI Tooling Visibility

⚡ Quick Summary

A Security Field Guide to AI Tooling Visibility by Monad Press provides a comprehensive security analysis of telemetry, native audit logs, and monitoring gaps across eight widely deployed AI developer tools and enterprise platforms: Claude Code, Claude Cowork, OpenAI Codex, Cursor, GitHub Copilot, Anthropic Compliance API, OpenAI API Platform, and Google Workspace Gemini.

The guide establishes an evidence boundary model categorizing AI logs into runtime OpenTelemetry evidence, control-plane audit events, and feature-invocation records. It details native schema fields, API pagination quirks, default logging configurations, retention limits, and concrete detection queries. Emphasizing a pipeline-first approach, the authors explain how to ingest, flatten nested OTLP records, enrich telemetry with identity and endpoint context before SIEM ingestion, and isolate sensitive prompt and tool bodies behind strict access and retention controls.

🧩 What's Covered

The field guide breaks down AI observability across two main operational categories and eight specific tooling environments:

• Local AI Agents and Developer Tools:
  • Claude Code OpenTelemetry: Evaluates metric exports, prompt capture (OTEL_LOG_USER_PROMPTS), command executions (tool_result, full_command), and Model Context Protocol (MCP) server connections, highlighting how to flatten raw nested OTLP payloads.
  • Claude Cowork OpenTelemetry: Details managed desktop agent telemetry across workspace paths, MCP invocations, and approval decisions, noting differences between metadata-only defaults and full content capture.
  • OpenAI Codex OpenTelemetry: Explores session-start sandbox policies (e.g., danger-full-access), approval drift, unredacted prompt exposures, and the necessity of endpoint telemetry for tracking CLI bypass flags.
  • Cursor Audit Logs: Outlines control-plane tracking via /teams/audit-logs—covering team roles, API key lifecycle, Privacy Mode toggles, and MCP server configurations—while identifying blind spots regarding local agent code execution.
  • GitHub Copilot Audit Logs: Analyzes enterprise audit streams (action:copilot) and cloud agent actions (actor:Copilot), content exclusion limitations, firewall allowlist updates, and seat lifecycle monitoring.
• Enterprise AI Platforms:
  • Anthropic Compliance Activity Feed: Explains monitoring for high-privilege operations, Compliance API key usage, and bulk workspace deletions across linked organizations.
  • OpenAI API Platform Audit Logs: Details governance detections mapped to MITRE ATT&CK, focusing on SCIM disabling, IP allowlist broadening, role escalation, BYOK tampering, and API call logging modifications.
  • Google Workspace Gemini Activity Logs: Examines gemini_in_workspace_apps reports, filtering active versus inactive event categories, and navigating Google Vault retention boundaries.
• Operationalizing Telemetry: Architecting pipeline control layers to normalize schemas, handle rate limits, redact prompt text, and enrich events with identity data before ingestion.

💡 Why it matters?

As autonomous AI agents and integrated enterprise assistants execute code, modify files, and access internal systems, security teams cannot rely on generic assumptions about vendor logging. Native AI logs vary dramatically: some capture runtime tool execution, others record only administrative policy changes, and some merely note feature invocations without prompts or results. This guide delivers the architectural clarity required to identify visibility blind spots, avoid treating control-plane logs as runtime detection feeds, and construct defensible logging architectures that protect sensitive prompt data.

❓ What's Missing

The guide focuses strictly on native logging capabilities emitted by vendor APIs and OpenTelemetry streams. It explicitly excludes deep endpoint runtime inspection (such as EDR process lineage or eBPF tracing), network NetFlow collection, and specialized prompt-injection defense layers or sidecar inspection tools. It also notes that telemetry schemas and API parameters evolve rapidly, requiring teams to continuously validate fields against their specific active tenant versions.

👥 Best For

Security engineers, SOC analysts, detection engineers, AI governance officers, and cloud security architects responsible for auditing, securing, and integrating AI developer agents and enterprise AI platforms into organizational SIEM and data pipelines.

📄 Source Details

• Title: A Security Field Guide to AI Tooling Visibility: The logs, blind spots, and security use cases across today’s most popular AI tools
• Authors: Darwin Salazar with Matt Jane, Curtis Redgate, Kenneth Kaye, and Valerie Worman
• Publisher: Monad Press
• Format: Technical Security Field Guide / Monograph (78 pages)

📝 Thanks to

Special thanks to Darwin Salazar, Matt Jane, Curtis Redgate, Kenneth Kaye, Valerie Worman, and the Monad research team for compiling this open and practical reference for securing modern AI tooling ecosystems.