CAIN-42 CAIN StudioCART · Module 2 of 5 · 15 min
Attack families against agents
Attack the words: direct and indirect prompt injection (through documents, web pages, tool returns), and goal hijack.
Attack the authority: delegation laundering (passing authority through another agent to hide its source), self-approval, token replay, privilege escalation through tool chaining.
Attack the boundaries: cross-tenant reads, homoglyph and path-traversal tricks, salami-sliced actions under per-action limits, and memory poisoning that persists into later tasks.
Key points
- Words: injection, indirect injection, goal hijack.
- Authority: laundering, self-approval, replay, escalation.
- Boundaries: tenants, homoglyphs, traversal, salami, memory.
In the CAIN-42 codebase
platform-gateway/adversarial_engine.py— Moat #6 attack engineplatform-gateway/cain_adversarial_suite.py— the adversarial suite
Practice check
Practice 1 · choose one
Indirect prompt injection arrives through…
Practice 2 · true or false