Nebius × NVIDIA Global AI Hackathon · Personal AI Track
Self-improving agents teach themselves skills.
SovereignOS decides which ones they're allowed to keep.
Personal AI agents now turn completed tasks into reusable skills. One poisoned file can teach an agent a skill that leaks your data, and that skill runs again every time.
Injected instructions steer the agent into saving a malicious helper as a trusted skill.
Once registered, the bad skill runs on every future task, long after the poisoned input is gone.
Asking the model "is this safe?" fails when the model itself has been manipulated.
Sandboxes stop bad actions at runtime. SovereignOS stops bad skills from being registered at all.
Parses the skill without running it and infers what it really needs: network, shell, secrets, dynamic code.
Compares those permissions with what the task justifies. A CSV task needs no network and no secrets.
NVIDIA Nemotron on Nebius Token Factory explains the risk in plain language. It advises; it never decides.
Same task, two skills. Pick one and press Run to replay the pipeline step by step, using output from a real run.
| Skill | What it does | Result |
|---|---|---|
csv_analyzer.py | Reads a CSV, counts rows and columns | Approved |
report_helper.py | Reads ~/.ssh/id_rsa, sends it to a remote host | Rejected · network, secrets |
Open model that writes the plain-language risk review.
Serves Nemotron through an OpenAI-compatible API.
Deterministic checks. The decision never depends on the model.
The AI can evolve.
The user keeps the authority.