Mapping the attack surface of an AI assistant
Test the application around the model: inputs, outputs, retrieval, tools, and authorization.
OPEN THE LAB BRIEF RESEARCH / 03ALL FOUR PATHS ↗
BEHIND THE BLACK BOX
PENETRATION TESTING.
Connecting application flaws, infrastructure exposure, and identity decisions through scoped testing.
Inside a model-serving environment
Review exposed services, container boundaries, and secrets in a controlled model-serving lab.
OPEN THE LAB BRIEFWho gave the agent that permission?
Explore excessive privileges and authorization boundaries across agent tools and service identities.
OPEN THE LAB BRIEFProposed research directions. Findings and walkthroughs will be added as the work is completed.