amal66 proposes a serious server-security upgrade for Mike

A closed, unmerged proposal would add practical safeguards around connected tools, confidential files and AI-generated legal answers.

The package treats legal AI as an application that must defend itself at several layers, not merely trust the model. It would:

block connectors from reaching internal systems or cloud credentials;
keep uploaded document text distinct from trusted workflow instructions;
require expiry on new document-download links while preserving old links during transition;
encrypt stored connector credentials and restrict which websites can call the service;
check quoted citations against the underlying documents without another AI call;
add automated dependency checks and reviewable security updates.

The proposal was closed without merging, so these protections are not part of the fork's delivered product. Still, it shows amal66 thinking through the operational risks that surface once an AI assistant can read client material and call external tools.