CAIN-42 CAIN StudioCCAS · Module 2 of 5 · 16 min
Threat modeling with the OWASP Agentic Top 10
The OWASP Top 10 for Agentic Applications for 2026 lists the most critical risks for autonomous agents, from agent goal hijack to tool misuse and the abuse of identity and privilege. It is the shared checklist most reviews now start from.
Threat-model each agent along its action paths: what can it reach, with whose authority, under what conditions? Goal hijack becomes a question about whether a hijacked plan can still execute a harmful action. Tool misuse becomes a question about argument validation and scope at the boundary.
Map each risk to a control that runs at execution time and to evidence that the control ran. A control that only exists in a policy document does not count.
Key points
- Use the OWASP Agentic Top 10 as the starting checklist.
- Threat-model action paths, not just prompts.
- Every risk needs a runtime control and evidence it ran.
In the CAIN-42 codebase
platform-gateway/adversarial_engine.py— attack families CAIN runs against itself, including goal-hijack and tool-misuse style cases
Practice check
Practice 1 · choose one
What does OWASP's 2026 agentic list cover?
Practice 2 · true or false