Menu ☰
aegis · · 3 min read

Pentest AI Agentic 2026: 5 Vulnerabilities Beyond OWASP Top 10 LLM

Three weeks of red-team engagements against agentic stacks this autumn, and the same pattern keeps surfacing: the OWASP Top 10 for LLM Applications catches almost none of what we actually exploit.

CAI Technology · Last reviewed: 10/16/2026
Clean editorial photo of three professionals at a conference table in a modern office; no text, no third-party logos, anatomy looks correct. Mood is somewhat somber/dim rather

Pentest AI Agentic 2026: 5 Vulnerabilities Beyond OWASP Top 10 LLM

Three weeks of red-team engagements against agentic stacks this autumn, and the same pattern keeps surfacing: the OWASP Top 10 for LLM Applications catches almost none of what we actually exploit. Prompt injection and insecure output handling are solved problems on paper. The money breaches sit elsewhere.

OWASP’s catalog assumes a single model behind an API. Agentic architectures run six-to-twelve autonomous nodes, each with its own tool belt and persistent memory. The attack surface moves from the model to the orchestration graph — exactly where traditional pentest playbooks go blind.

Where OWASP stops

OWASP Top 10 for LLM v1.1 still indexes intake-layer threats: prompt injection, training-data poisoning, model DoS. Agentic systems add runtime concerns — inter-agent trust, delegated tool authority, long-horizon working memory. The ENISA Threat Landscape 2025 flags autonomous AI agents as an emerging category requiring new control taxonomies, and the NIST AI RMF Generative AI Profile separates single-model from compound-system risk.

The five new attack classes

1. Tool-definition poisoning. An attacker who controls one MCP server rewrites a tool’s description, not its code. The LLM reads the description, decides to invoke, and ships credentials to the attacker’s endpoint. The model behaved correctly — OWASP cannot see this.

2. Working-memory contamination. Long-horizon agents persist state in vector stores and scratchpads. A payload injected on run 47 fires on run 203. Pentests scoped to session boundaries miss it entirely.

3. Multi-agent collusion. When two sub-agents hand each other inputs, a backdoor split across both never fires in isolation. See the taxonomy of six failure clusters in LLM agent failure taxonomy.

4. Delegation privilege escalation. Agent A has read-only access, agent B has write. A asks B to “update the shared doc” on its behalf. The system logs B as actor. The authorization model collapses silently.

5. Reasoning-chain hijack. The attacker contaminates not the retrieval corpus but the chain-of-thought stream. The agent reasons toward the wrong goal using legitimate data.

# agent_runtime audit log fragment
2026-10-14T09:22:11Z orchestrator: agent=planner_v3 tool_called=notion.update_page
2026-10-14T09:22:11Z orchestrator: delegated_from=reader_agent (scope=read-only)
2026-10-14T09:22:12Z policy_engine: VIOLATION privilege_lift write_via_delegation
flowchart TD A[User prompt] --> B[Planner agent] B --> C{Tool metadata trusted?} C -->|no| D[Reject + alert SOC] C -->|yes| E[Delegate to worker] E --> F{Delegated scope ⊆ caller scope?} F -->|no| G[Privilege escalation] F -->|yes| H[Execute + log provenance] classDef bad fill:#fee2e2,stroke:#ef4444 classDef good fill:#dcfce7,stroke:#10b981 class D,G bad class H good

Methodology shift

Pentesting an agentic deployment is a graph exercise, not a prompt exercise. Engagements aligned with EU AI Act Article 15 now map every tool node, memory boundary, and delegation edge before any payload. Supply-chain pressure from NIS2 Directive (EU) 2022/2555 raises the stakes: compromised agents are reportable incidents. The complementary defender-side view sits in where LLM cyber agents break.

OWASP v2 will likely absorb three of the five. Delegation escalation and reasoning-chain hijack need primitives OWASP’s framing cannot host. CAI’s AEGIS red-team pack ships scenario modules for all five; the discovery phase is where the real work sits, not the exploitation. Talk to the AEGIS team before your next audit cycle.

Read further

We start with a 30-minute conversation.

Free AI-readiness audit for companies with 50+ employees. We reply within 24 hours.