Securing the Company Brain: A Wiki-First Architecture
AI Engineergo watch the original →
the gist
To prevent data leaks in AI company brains, move away from autonomous agent memory toward a human-in-the-loop wiki model where agents suggest changes, users verify them, and access is strictly scoped per file.
The Problem with Autonomous Memory
Building a "company brain"—an AI system that learns from internal data—is often stalled by security fears. Organizations, particularly in regulated industries like banking, cannot risk agents leaking sensitive information (e.g., compensation data) to unauthorized users. Current approaches often fail because they either rely on developers manually writing shared skills in GitHub (which rarely happens) or create isolated silos (e.g., per-channel AI memory) that prevent cross-functional knowledge sharing.
The Wiki-First Architecture
Instead of treating the company brain as a black-box database, the solution is to model it as a centralized wiki of linked markdown files. This approach enforces two critical rules: all context must live in one shared space, and every single change must be attributed to a human. By requiring a human to accept or reject an agent's proposed knowledge update, the system maintains accountability. If a leak occurs, there is a clear audit trail identifying who approved the inclusion of that specific information.
Scoping and Credential Injection
Security is managed through granular access control lists (ACLs) applied at the file level. When an agent performs a task, it does not operate with a global "god-mode" credential. Instead, it acts as a proxy for the specific user currently interacting with it, injecting that user's credentials at the HTTP and SQL layers. This ensures that the agent can only read or write to parts of the wiki that the human user is authorized to access, effectively preventing privilege escalation.
Multiplayer Knowledge Creation
In complex scenarios like incident management, knowledge is often generated through debate between engineers. The system should capture these interactions as they happen. By allowing multiple users to debug an issue while the agent observes and proposes "learnings" (e.g., "use an equals query instead of a like query"), the team builds the company brain as a byproduct of their actual work. This transforms the brain from a static repository into a living, growing asset that reflects the collective expertise of the organization.