保存先:
| 第一著者: | |
|---|---|
| フォーマット: | Recurso digital |
| 言語: | |
| 出版事項: |
Zenodo
2026
|
| 主題: | |
| オンライン･アクセス: | https://doi.org/10.5281/zenodo.20012329 |
| タグ: |
タグ追加
タグなし, このレコードへの初めてのタグを付けませんか!
|
目次:
- <p class="MsoNormal"><span>Autonomous AI agents capable of reasoning, planning, and interacting with external systems are increasingly integrated into enterprise infrastructures. In parallel, recent industry analyses have highlighted the growing sophistication of automated agents, including their ability to bypass traditional detection mechanisms and exploit application logic.</span></p> <p class="MsoNormal"><span>These developments are typically addressed from distinct perspectives: external malicious automation and internal AI agent security. Taken together, however, they suggest a deeper structural issue. This paper argues that both domains converge toward a fundamental problem: untrusted decision sources interacting with trusted systems.</span></p> <p class="MsoNormal"><span>Building upon prior work introducing governance-centered control of execution authority, this paper proposes a unified security model in which execution authority is decoupled from reasoning processes and enforced through an independent Policy Root of Trust. </span><span>By treating all protected operations as action requests requiring independent authorization</span><span>, this approach complements existing detection mechanisms and </span><span>provides enforceable execution-time controls</span><span>.</span></p>