A pragmatic threat model for AI coding agents, with controls you can ship today

Short summary

OWASP's 2025 agentic applications threat model identifies nine critical failure modes including prompt injection, overprivileged tools, data leakage, and hallucination. This practical guide translates each into concrete, immediately shippable controls using Akmon—policy profiles, tool surface restrictions, structured outputs, and version pinning. Each control includes code examples, real incident patterns, and explicit trade-offs.

• •OWASP Top 10 for Agentic Apps provides shared vocabulary for AI agent failure modes and incident patterns
• •Each threat maps to a concrete control with Akmon configuration, CLI flags, and runnable code examples
• •Trade-offs are explicit—stricter policies add friction but prevent categories of incidents

Generated with AI, which can make mistakes.