Excessive Agency
The assistant needed to read invoices. Its credential could also change any supplier’s bank account, and nothing asked before it did.
-
-
Prompt Injection and Untrusted Content Boundaries
A hostile sentence in a note, passage or tool result can steer a model. Track where text came from, bound what a steered model can do, and test defenses against adaptive attacks.
-
Privacy, Data Leakage, and Output Controls
A request leaves copies in logs, caches, provider storage and test sets. This post builds redaction, permission rechecks, output validation and a deletion inventory, and tests them together.