What does human control over AI actually mean?
It means deciding what an assistant may read, change, send or spend—and being able to withdraw that permission. A useful assistant can work within a limited task without asking about every small step. It should return to you when the task needs more access, a new destination or a larger budget.
For example, permission to draft an email is not permission to send it. Access to one project folder is not access to your whole computer. Knowing your preferences is not authority to make commitments in your name.
More capable intelligence should not automatically receive more power over your life.
Can an AI agent act without my permission?
That depends on the software around the model and the access it has been given. A model connected to broad credentials or powerful tools may be able to act without a fresh decision from you. A prompt saying “always ask first” is not the same as an enforced permission boundary.
AUKORA’s Golden Boundary proposes separating the software that requests an action from the authority that permits it. The permission must cover the actual operation, target and limits. The model must not be able to bypass that route or approve a wider grant for itself. Establishing that separation across every reachable route remains unfinished.
Can an AI agent read or share my private data?
The key questions are which data the system can access, where it sends that data, and what each service keeps. Privacy can be lost when a file is uploaded for advice, before the assistant takes any later action. A provider’s retention rules and the actual data path both matter.
Encrypting a connection protects it in transit; ordinary remote inference still gives the receiving service access to the content it processes. A local model changes that arrangement, but local software can still leak information through plugins, network access or inherited credentials. “Local” and “open source” are not automatic guarantees.
What happens to my memory when I change AI providers?
AUKORA’s goal is that useful history belongs to the person rather than to one model. Kira is designed to keep source records distinct from summaries and AI inferences, so a replacement model can work from retained context without turning yesterday’s guess into a fact.
But an export button alone does not establish independence. The records must be usable elsewhere, identity must survive a move, and old permissions must not broaden during migration. This portability is an architectural goal to test, not a claim that seamless migration between all providers already works.
What is AUKORA researching?
AUKORA Sovereign is an open-source personal AI node, with its authority written in Rust, exploring human sovereignty for personal AI agents. Its focus is the relationship between identity and permission (Aumlok), memory and provenance (Kira), and inspectable records (Aura). The aim is an assistant you can use, question, limit and eventually replace without losing your digital continuity.
The builders record what a disposable test node did, with a software passkey and a fake model key, commands and output included. A stranger’s verifier checks signed records offline from an empty folder, and states its own limits. That gives readers specific claims to inspect. It does not establish the safety of the whole system, the identity of the person who approved, or the truth of every recorded event.
Has AUKORA solved AI safety?
No. Nothing is live: no owner has claimed a node, and all eight milestones are still open. Recovery after losing every device, sealed memory, exit drills and node-to-node messaging are not built yet. On a rented machine the host can read a running node, and what the agent reads can still leave in its prompts.
The Golden Boundary addresses authority and continuity. It does not prove that an AI is truthful, harmless, aligned or free from persuasive influence. A signature can record a decision without proving understanding or freedom from coercion. This is work to examine—not a reason to hand a prototype sensitive data or unrestricted access.
What should I ask before giving an agent access?
Ask what it can read and change; who holds the credentials; which actions require a separate approval; how time, spending and resource limits are enforced; and how you can revoke access. Ask what gets sent to remote providers and what survives in logs or backups.
Then ask for evidence: can it exceed a grant, reuse a spent permission, rewrite a recorded outcome or prevent you from leaving? A useful answer names the tested version, the observed result and the remaining assumptions. A reassuring animation or a confident explanation is not that evidence.
Where can I read the research and check it?
The human-readable Golden Boundary explains the full argument. The canonical technical paper states the mechanisms and assumptions. Current claims and limits marks every claim LIVE, RECORDED, TESTED, SOURCE or NOT YET, with its evidence.
For concrete evidence, read the test node’s record and run the stranger’s verifier. For the implementation map, read the Sovereign README. Questions and careful criticism are welcome through the contribution guide.