Runtime Attribution Framework: An External C-ABI and PKI-Based Zero-Trust Infrastructure for Non-Repudiable Execution Governance in Multi-Agent Systems (中文對譯:智能體運行期歸責框架:多智能體系統中基於外部 C-ABI 與 PKI 零信任的不可否認性執行治理基礎設施)

[English] As autonomous AI agents scale across enterprise workflows, binding execution-layer privileges to specific agent identities remains a critical unsolved challenge. Existing AI governance solutions face severe architectural blind spots: semantic-layer guardrails (such as NVIDIA Nemo Guardrails) are vulnerable to jailbreaks, while traditional OS sandboxes (such as eBPF) suffer from "context-blindness." To resolve this "Attribution Gap," we present the Runtime Attribution Framework: an external C-ABI and PKI-based zero-trust infrastructure for non-repudiable execution governance in Multi-Agent Systems. By deploying enforcement gates directly at the binary FFI / C-ABI boundary, the framework cryptographically binds agent identity to allowed capability bitmaps, generating immutable, non-repudiable audit logs signed via Ed25519. [中文摘要] 隨著自主 AI Agent 的普及,將執行層權限與 Agent 身份進行強綁定已成為零信任架構中亟待解決的挑戰。我們提出了「智能體運行期歸責框架」。透過部署於二進位 FFI / C-ABI 邊界,將 Agent 身份與其能力點陣圖(BEC)進行密碼學綁定,並於 GuardVM 內以低於 1 微秒的 O(1)