LLM 曝露監視のオープンな記録パターン：ギャップ分析とスキーマ提案

AI agents access sensitive data across heterogeneous platforms, yet no open specification defines the exposure event semantics needed to make this access auditable. This paper proposes open recording patterns, an exposure event schema extending OCSF, CEF, and OpenTelemetry GenAI with context window events, attestation source fields, and agent attribution semantics. The schema defines what platforms report so that monitoring infrastructure can consume exposure data in a standardized format, analogous to XBRL for financial reporting. We define the LLM Exposure Surface (Read, Write, API, Context), identify five exposure-specific threats (ET1–ET5, mapped to STRIDE and MITRE ATLAS), and classify 15 platforms into three recording depth tiers, finding that only 1 of 13 native platforms (8%) provides real-time exposure event channels. To illustrate the cost of the current gap, we show that without platform-emitted access labels, monitoring precision is structurally limited by the access ratio (0.14–0.53 across configurations, N = 50 runs each); with schema adoption, simulated precision rises to ≥0.95, demonstrating the concrete benefit of standardized exposure semantics. A schema validation test (N = 50 live GitHub issues, each producing create and close webhook events) confirms correct end-to-end operation. The paper contributes a concrete schema specification for community evaluation and adoption, a systematic gap analysis of the current platform landscape, and a quantified case for standardization.