The Enterprise AI Brief | Issue 13
Inside This Issue
The Threat Room
When "Read Only" in the AI Connector Does Not Mean Read Only in the Database
A database connector says “read only.” Then a query it was supposed to stop gets through. A newly disclosed AWS vulnerability shows why the permissions behind an AI connector can matter more than the label in front of it.
→ Read the full articleThe Operations Room
New Models Arrive. Old Models Keep Consuming Capacity.
T-Tech had more than 200 internal applications sharing a finite GPU pool, and new model generations were arriving faster than old ones could be retired. Its solution to model sprawl was unexpectedly literal: build one more model, this time around the production behaviors that had prevented consolidation in the first place.
→ Read the full articleThe Engineering Room
The Agent Retrieved the Evidence and Then Failed to Use It
An AI agent can call the right tool, retrieve the evidence it needs, and still fail to use what came back. New research finds that agents with nearly identical accuracy can behave very differently behind the score, making the gap between a successful tool call and successful reasoning the harder thing to measure.
→ Read the full articleThe Governance Room
Sharing an Agent's Capability Without Sharing the Company's Playbook
An agent can use another company's expertise without receiving the prompt or code behind it. Skill-as-API shows how that boundary can work, but its audit trail exposes a harder governance problem: proving who had access, what they invoked, and which private procedure actually produced the result.
→ Read the full article