SAFETY & TRUST
An agent that can see a screen or use a camera is, functionally, remote access to a stranger’s device. That makes safety a product requirement, not a later phase, and it makes honesty about our progress part of the deal.
WHAT IS TRUE TODAY
THE NON-NEGOTIABLES
Every access scope (viewing the screen, controlling the device, using the camera) is a separate opt-in, limited to a single session, and can be withdrawn at any time. There is no silent or standing access.
A persistent indicator shows that the agent is active and what it is doing right now. Stopping the session ends access immediately.
The agent receives the minimum access the specific case needs, never blanket device control.
Deleting files, reinstalling an OS or factory-resetting all require explicit customer confirmation, with rollback where it’s possible.
Everything the agent does during a session is logged, so a customer or a company reviewer can see exactly what happened and why.
Retention limits, what is kept versus discarded after a case closes, and how learned fixes are generalised without leaking private data between customers are defined before launch, not after.
An agent with device access is a target. We design for prompt injection from on-screen content, cloned support pages and social engineering from the start.
DESIGNING AGAINST ABUSE
These are the threats the design has to answer before any real device is touched.
A web page, document or error dialog can contain text written to hijack an AI agent. Everything the agent reads from a session is treated as untrusted data and never as an instruction. Actions come only from the plan the customer has approved.
An attacker could imitate a company’s support page to get a customer to grant access. Sessions must be bound to a verified company origin so a lookalike page can’t start one.
The agent must never ask for passwords or one-time codes, and must resist requests to widen its own scope. Customers are shown plainly what is being asked and why.
Scope can’t be widened silently. Any extra permission is a new, visible request that the customer can refuse without losing the help already given.
THE REGULATORY LANDSCAPE
This is a map of obligations we expect to face, not a statement of compliance.
| Framework | Why it applies | What it requires | Our status |
|---|---|---|---|
| GDPR / UK GDPR | When personal data of people in the EU or UK is processed, which screen and camera access almost certainly involves. | A lawful basis, data minimisation, purpose and retention limits, deletion rights and, in all likelihood, a data protection impact assessment before launch. | Not started. Design principles only. |
| CCPA / CPRA | When serving California residents’ personal information. | Notice at collection, rights to know and delete, and no sale or sharing of personal information. | Not started. |
| EU AI Act | Transparency duties for AI systems that interact directly with people. | People must be told they are interacting with an AI system. | Requirement flagged for the architecture. |
| SOC 2 | Enterprise buyers will ask for it before trusting a vendor with device access. | Documented security controls, operated and evidenced over an audit period. | Not certified. No audit has been started. |
| HIPAA | Only if health-adjacent customers are served. | Business associate agreements and specific safeguards for protected health information. | Out of scope unless a customer segment requires it. |
THIS WEBSITE
Only what you submit in the early-access form, deleted automatically after 90 days. No cookies for visitors, no analytics, no advertising trackers, and no request for screen, camera or device access. Read the privacy notice for the details.
Tell us you’re interested. No access to any product is granted, and you can delete your details at any time.