Slide 27 of 27
CompleteSlide 27
Complete
LLM06:2025 Complete
You’ve covered all 5 parts, read the real CVEs, and tested yourself.
What Excessive Agency is — in plain English and in OWASP’s definition
The three root causes: excessive functionality, permissions, and autonomy
Why LLM06 is an architectural failure, not a coding bug
How LLM06 amplifies every other OWASP LLM risk into real-world action
CVE-2025-53773 (GitHub Copilot, Aug 2025) — indirect injection → YOLO mode → malware → C2
CVE-2025-54136 (MCPoison, Apr 2025) — tool description poisoning, 60–72% success rate
Slack AI exfiltration (Aug 2024) — overpowered data access exploited via indirect injection
4 attack patterns: overpowered tools, excessive permissions, MCP poisoning, indirect injection
All 4 OWASP official scenarios: email agent, document agent, shell plugin, cascading agents
7 mitigations: least-priv functionality, least-priv permissions, human-in-the-loop, external authz, tool allowlisting, sandboxing, audit logging
Review from beginningNext: LLM07 →
📄 View all sources for this lesson