Legal AI Is Broken at the Architecture Level — Here’s What to Fix

Building AI for legal or compliance teams is a different threat model than most devs are used to. It’s not just external attackers you’re defending against. It’s regulators who can demand documented evidence of your data governance. Auditors who test whether your stated controls exist in your running systems. Opposing counsel who can subpoena records of how AI processed sensitive client data. The standard “we encrypt at rest, we’re behind auth” answer doesn’t hold up in this environment. Here’s what actually needs to change — and the resources that go deeper on each piece. The Four Things Most Legal AI Stacks Get Wrong No anonymization pipeline . Raw PII — names, account numbers, case notes — flows directly into model training. Tokenization and aggregation should be first-class steps in ingestion, not afterthoughts. Partial encryption Production is locked down. Dev and staging are not. A misconfigured dev bucket has the same regulatory consequence as a production breach if it contains