Приложение А — Матрица сопоставления безопасности OWASP Agentic AI
Кросс-маппинг ASI Top 10, LLM Top 10, Agentic AI Threats & Mitigations и AIVSS Core Risks
| ASI ID / Название | OWASP LLM Top 10 (2025) | Agentic AI Threats & Mitigations | AIVSS Core Risk Alignment |
|---|---|---|---|
| ASI 01 — Agent Goal Hijack | LLM01:2025 Prompt Injection · LLM06:2025 Excessive Agency | T6 Goal Manipulation · T7 Misaligned & Deceptive Behaviors | Agent Goal & Instruction Manipulation |
| ASI 02 — Tool Misuse & Exploitation | LLM06:2025 Excessive Agency | T2 Tool Misuse · T4 Resource Overload · T16 Insecure Inter-Agent Protocol Abuse | Agentic AI Tool Misuse |
| ASI 03 — Identity & Privilege Abuse | LLM01:2025 Prompt Injection · LLM06:2025 Excessive Agency · LLM02:2025 Sensitive Info Disclosure | T3 Privilege Compromise | Agent Access Control Violation |
| ASI 04 — Agentic Supply Chain Vulnerabilities | LLM03:2025 Supply Chain Vulnerabilities | T17 Supply Chain Compromise · T2 Tool Misuse · T11 Unexpected RCE · T12 Agent Comm Poisoning · T13 Rogue Agent · T16 Insecure Inter-Agent Protocol Abuse | Agent Supply Chain & Dependency Attacks |
| ASI 05 — Unexpected Code Execution (RCE) | LLM01:2025 Prompt Injection · LLM05 Improper Output Handling | T11 Unexpected RCE & Code Attacks | Insecure Agent Critical Systems Interaction |
| ASI 06 — Memory & Context Poisoning | LLM01:2025 Prompt Injection · LLM04:2025 Data & Model Poisoning · LLM08:2025 Vector & Embedding Weaknesses | T1 Memory Poisoning · T4 Memory Overload · T6 Broken Goals · T12 Shared Memory Poisoning | Memory Use & Contextual Awareness |
| ASI 07 — Insecure Inter-Agent Communication | LLM02:2025 Sensitive Information Disclosure · LLM06:2025 Excessive Agency | T12 Agent Communication Poisoning · T16 Insecure Inter-Agent Protocol Abuse | Agent Memory & Context Manipulation |
| ASI 08 — Cascading Failures | LLM01:2025 Prompt Injection · LLM04 Data & Model Poisoning · LLM06:2025 Excessive Agency | T5 Cascading Hallucination Attacks · T8 Repudiation & Untraceability | Agent Cascading Failures |
| ASI 09 — Human-Agent Trust Exploitation | LLM01:2025 Prompt Injection · LLM05:2025 Improper Output Handling · LLM06:2025 Excessive Agency · LLM09 Misinformation | T7 Misaligned & Deceptive Behaviors · T8 Repudiation & Untraceability · T10 Overwhelming Human in the Loop | Agent Untraceability / Human Manipulation |
| ASI 10 — Rogue Agents | LLM02:2025 Sensitive Information Disclosure · LLM09:2025 Misinformation | T13 Rogue Agents in Multi-Agent Systems · T14 Human Attacks on Multi-Agent Systems · T15 Human Manipulation | Behavioral Integrity (BI) · Operational Security (OS) · Compliance Violations (CV) |
Примечания
- LLM Top 10 alignment: отражает корневые уязвимости LLM, расширенные на агентные системы.
- Agentic Threats & Mitigations (T1–T17): представляют детальные пути атак, на которые ссылается ASI-фреймворк.
- AIVSS Core Risks: сопоставляются с количественными категориями оценки для приоритизации и ранжирования критичности (например, BI, OS, CV).
- Crossover insight: записи ASI часто объединяют несколько записей LLM — например, ASI01 сочетает LLM01:2025 (prompt) с LLM06 (autonomy) — отражая, как агентная автономность усиливает риски на уровне модели.
genai.owasp.org