Neutralizing Prompt Injection and LLM Hallucinations: The Deterministic Sandboxing of VajraAgent and VajraClaw(防護指令污染與幻覺危害的DROS:VajraAgent 與VajraClaw 的執行治理基礎設施)

Large Language Models (LLMs) inherently possess black-box flaws, including unpredictable hallucinations and extreme vulnerability to Prompt Injection. When these models are encapsulated as AI Agents with autonomous decisionmaking capabilities and deployed in enterprise environments, their uncontrollability can lead to devastating unauthorized access and data breaches. Rather than attempting to eradicate hallucinations from within the model (which has proven to be extremely difficult and unreliable under current neural network architectures), this paper proposes an external execution governance infrastructure based on DROS (Deterministic Runtime Operating System)—the Vajra System. Through the decoupled VajraClaw (Vajra Claw) and the VajraAgent (Vajra Agent) equipped with a deterministic Finite State Machine (FSM), weimplement strict “Semantic & Physical Path Isolation” at the system level. In a batch evaluation containing 1,000 malicious prompt injection and privilege escalation test samples, the VajraAgent state machine successfully filtered and blocked98.5% of for- matting anomalies and explicit overstepping. The remaining 1.5% of sophisticated forged commands that bypassed the semantic layer were 100% physically circuitbroken at the C- FFI binary execution layer by VajraClaw’s path and signature validation, achieving system-level zero overstepping. While providing military-grade security, this architecture introduces an extremely low latency overhead of under 1ms and an edge-end SDK memory footprint of under 10MB, laying a cornerstone for the secure, traceable “AgentWeb” era.-----------------------------------------------------------------------------------------------------------------------------------------大型語言模型(LLM) 先天存在不可預測的幻覺(Hallucinations) 與極易遭受提示詞注入(Prompt Injection) 的黑盒子缺陷。當這些模型被封裝成具備自主決策能力的AI Agent 並部署於企業環境時,其不可控性將導致毀滅性的越權行為與資料外洩。與其試圖從模型內部根除幻覺(這在目前的神經網路架構下已被證明是極度困難且不可靠的),本文提出了一套基於DROS (Deterministic Runtime Operating System) 的外部執行治理基礎設施——Vajra 體系。透過解耦的V ajraClaw (金剛爪) 與具備確定性有限狀態機(FSM) 的VajraAgent (金剛智能體),我們在系統層級實作了嚴格的「語義與物理路徑雙重隔離(Semantic & Physical Path Isolation) 」。在包含1,000 個惡意指令注入與越權測試樣本的批量評估中,VajraAgent 的狀態機成功過濾並攔截了98.5% 的格式異常與顯性越權,而剩餘1.5% 成功繞過語意層的精緻偽造指令,則在C-FFI 二進位執行層被VajraClaw的路徑與簽章校驗100% 物理熔斷,實現了系統級的零越權。本架構在提供軍規級安全性的同時,帶來低於1ms 的極低延遲開銷與小於10MB 的邊緣端SDK 記憶體佔用,為安全、可溯源的「AgentWeb」世代奠定了基石。