رؤى / تقرير بحثي
أمّن الصلاحيات الكامنة وراء الإجابة
تصميم مسارات ذكاء اصطناعي مفيدة ومحدودة الأثر حتى عند التعرّض للهجوم
توليف للمصادر الأولية وتحليل mayden. المنهج والحدود موضحان.
الملخص التنفيذي
- حدّد ما تستطيع الخدمة قراءته واقتراحه وتنفيذه، وافصل الإذن عن تفسير النموذج.
- اختبر إنجاز العمل المشروع ومقاومة الهجمات كلّاً على حدة.
- استخدم صلاحيات محددة وموافقات مرتبطة بالفعل وخطة تعافٍ مجرّبة.
اقرأ التقرير الكامل بالإنجليزية
يتوفر الملخص التنفيذي بالعربية. يتضمن التقرير الإنجليزي التحليل الكامل والرسوم التوضيحية والأسئلة القيادية والمراجع.
قراءة التقرير بالإنجليزيةتنزيل التقريرمنهج البحث وحدوده
كيفية قراءة هذا البحث
تفاصيل المنهج والمراجع أدناه باللغة الإنجليزية.
- Approach
- Narrative synthesis of eight opened primary research and official or community technical sources, accessed 10 October 2026; original mayden authority framework.
- Source selection
- Indirect prompt injection, tool-agent evaluation, least privilege, zero trust, lifecycle development and adversarial threat taxonomy.
- Limitations
- No penetration test of a live system, proprietary incidents or quantitative meta-analysis. Historical demonstrations do not establish current attack rates. Guidance does not certify an implementation. The procurement case is fictional.
المراجع / 8 مصادر
الأدلة التي يستند إليها التحليل
- [1]Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection ↗
Greshake et al. · 2023-05-05, arXiv v2
Primary security research · Abstract
Historical demonstrations of indirect prompt injection in specified systems; not a measurement of current attack incidence or proof that all systems share the same vulnerability.
- [2]LLM01:2025 Prompt Injection ↗
OWASP GenAI Security Project · 2025 edition; accessed 2026-10-10
Community security guidance · Prevention and mitigation
Community mitigation guidance; instruction separation, limited permissions and human approval reduce particular risks without guaranteeing prompt-injection elimination.
- [3]AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents ↗
Debenedetti et al.; NeurIPS · 2024-11-24, arXiv v3
Peer-reviewed benchmark research · Abstract
Research benchmark with 97 tasks and 629 security cases; simulated attacks do not establish real-world incident frequency or complete threat coverage.
- [4]SP 800-207: Zero Trust Architecture ↗
NIST · 2020-08
Technical guidance · Abstract
General zero-trust guidance rather than an AI-specific control certification; its use at AI tool boundaries is mayden synthesis.
- [5]Guidelines for secure AI system development ↗
UK National Cyber Security Centre and international partners · 2023-11-27, version 1.0
Official technical guidance · Four lifecycle areas
Official secure-development guidance organized around four lifecycle areas; not a certification or proof of prevention effectiveness.
- [6]SP 800-218A: Secure Software Development Practices for Generative AI and Dual-Use Foundation Models ↗
NIST · 2024-07
Technical guidance · Abstract
A secure-development profile for producers and acquirers of generative AI and foundation models; conformance with a process does not establish absence of vulnerabilities.
- [7]AI 100-2 E2025: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations ↗
NIST · 2025-03; corrected PDF April 2025
Technical taxonomy · Abstract; publication planning notes
A taxonomy of attacks and mitigations, not an incidence dataset or security certification. The March 2025 publication has later correction/planning notes on the official page.
- [8]Summary of Comments on the Concept Paper — NCCoE Agentic AI Identity and Authorization ↗
NIST NCCoE · 2026-09-29 release; living resource accessed 2026-10-10
Official consultation synthesis · Delegation and Accountability
Commenters' views; not a final standard or quantified control-effectiveness study.
