Another employee's HR data
An ordinary employee obtains salary or evaluation information through a broadly authorized agent.
AI Agent RisksGreater autonomy creates
Greater autonomy creates
more boundaries to cross.
The material risk is not only an incorrect answer. It is an unauthorized data access, action, transmission, or tool sequence that actually succeeds.
We assess authorization, data, actions, tools, destinations, and auditability.
Alternative data paths
After denial by the primary database, the agent searches a DWH, file repository, RAG index, or email.
Unauthorized operations
A read-only user causes an update, export, external transmission, or deletion through an agent tool.
Instructions inside content
The agent follows malicious instructions embedded in retrieved documents or web pages.
Inconsistent decisions
Equivalent requests receive different authorization outcomes after rephrasing or multi-turn prompting.
Missing evidence
There is no reliable record of data sources, tool calls, denial decisions, or post-denial attempts.
Next
Scope & Exclusions →Define what is and is not tested.
The diagnostic scope, exclusions, data, accounts, and stop conditions are agreed before testing.