SUPER LABS / RESEARCH NOTE

Access is not authority.

Possession answers what a system can reach. Authority answers what it may do, for whom, and why.

SUPER Labs · 12 Oct 2026 · Original analysis

  1. Requested action

  2. Purpose + consent

  3. Permission boundary

  4. Human-authorized action

Enterprise permissions were largely designed around people working inside applications. Read access was a tolerable proxy for what could happen next because the application, role and workflow constrained the action. Agents dissolve that boundary. They can retrieve from one system, transform in another and disclose through a third. Possession of data no longer describes the authority of the resulting action. A useful control model therefore has to represent not only who can see a record, but which purpose, recipient, transformation and execution path are permitted now.

Consider a partner-performance export. A data owner has approved a quarterly report containing aggregate service levels for a named recipient. The underlying warehouse also contains customer names, support transcripts and row-level transaction history. An agent can technically retrieve all of it. The failure is not a dramatic intrusion: it is a plausible request—‘send the partner everything behind this chart’—executed through credentials that work. If the control checks only database access, the raw disclosure passes. The authority boundary should instead allow the aggregate report, deny personal records and hold an unrecognized recipient for review.

Our thesis is that data rights should travel with operational context. Within Super, Super Data would preserve provenance, freshness, sensitivity, consent, purpose and permitted use without requiring every source to be centralized. SuperEthos would evaluate the declared action against human-approved objectives and evidence. Super Control would enforce the resulting decision at a covered export or tool boundary. A hard denial remains deterministic: model reasoning can supply evidence or identify ambiguity, but it cannot negotiate away a prohibition.

Execution timing matters. An approval should bind a specific action hash, scope, recipient and expiry rather than create a reusable permission token. The covered executor should recheck material conditions immediately before release. If the partner changes, consent is revoked or the report now contains a restricted field, the previous approval should no longer authorize execution. A receipt should identify the principal, policy version, evidence and result so that an operator can reconstruct the decision without claiming access to a model’s private reasoning.

This design is intentionally concrete. It controls only implemented connectors and covered paths; an uncontrolled export remains uncontrolled, and a disclosure that already occurred cannot be undone. Those limits make the research more useful, not less. The claim to earn is not universal safety. It is that a clearly bounded system can make the difference between access and authority inspectable, enforceable and measurable in one consequential workflow.