Notifications
Clear all
Benchmarks and Evaluation Methodologies
1
Posts
1
Users
0
Reactions
14
Views
Topic starter
June 22, 2026 12:12 pm
I'm trying to learn red teaming for AI agents. Every threat model I see is too vague. "Attacker could do bad things." Not useful.
So I built a template. It's for regulated environments like finance or healthcare. It forces you to list specific assets (PII, database credentials, internal API keys), then map how an agent could exfiltrate them via function calls or memory. It has a table for attack paths: "1. Inject into system prompt to add a new function. 2. Use that function to write data to a cloud storage bucket the agent can access." I want to know if this is a good starting point for designing attacks. What is missing?