Security & privacy
Powerful AI requires stronger boundaries.
Security execution map
Model proposal → independent control → scoped action
Selected request · illustrative scenario
Move 248 files within an approved Documents folder.
Independent controls evaluate the proposed plan against granted scope, policy and risk. The model cannot override this boundary.
04 Permission Decision
Explicit approval needed
Awaiting user approval
The folder boundary is valid, but moving 248 files changes its structure. The example waits for the user to review the plan. Nothing executes while approval is pending.
Execution is locked
User review is the next step. No execution or result verification yet.
Ten principles
How the boundaries are drawn.
- 01In Development
Independent permission enforcement
The model proposes actions; a separate policy engine decides. The model cannot grant itself access.
- 02
AI Intent Compiler research
Compiling natural-language goals into a restricted, typed action plan that can be checked before execution.
- 03In Development
Tool-execution boundaries
Each tool runs with explicit inputs, outputs and limits; nothing is implicit.
- 04In Development
Least-privilege access
Permissions are scoped to specific resources and expire. Broad access is the exception, not the default.
- 05Planned
Privacy-aware cloud routing
When cloud models are used, policies decide what may leave the device. Choosing cloud means data is sent to that provider.
- 06
Untrusted-content handling
Web pages, emails and documents are treated as data, never as instructions with authority.
- 07
Prompt-injection defence
Layered mitigations reduce risk. No approach today fully prevents prompt injection, and we don't claim otherwise.
- 08Planned
Action verification
After execution, outcomes are checked against the approved plan.
- 09Planned
Reversible operations
Supported changes can be undone. Not every action is reversible, and the interface says so before you approve.
- 10Planned
Auditability and user control
A readable history of proposals, decisions and actions, owned and visible to you.
Scoped identity
Agents carry a passport, not a master key.
AI Action Passport
NV-PASS · 7F3A-21C9
- Agent Identity
- file-organiser · v0.1
- Allowed Capabilities
- list, create-folder, move
- Resource Scope
- ~/Documents (no subfolder escape)
- Permission Expiry
- 15 minutes after approval
- Risk Level
- Medium (approval required)
- Execution Status
- Awaiting your approval
What we don't claim
- No security certifications have been obtained yet.
- No independent audits or penetration tests have been published.
- No system is immune to prompt injection or attack.
- Cloud mode is not offline; data goes to the chosen provider.
Found a vulnerability? Please report it responsibly via our contact page.