We use cookies to keep the site working, learn how it's used, and improve what we ship next. Accept everything, reject the non-essential, or pick category by category. See our privacy policy.
Security model
The AI isn't in charge of your device. You are.
Giving an agent access to the physical world should never mean handing it unrestricted control.
Ask first.
Sensitive capabilities require an explicit, scoped user authorization.
Allow only what's needed.
Short-lived authorizations bind to exact action parameters rather than broad future access.
Reject unsupported actions.
A deny-by-default policy blocks unsupported requests before execution. Model output cannot trigger shell, ADB, arbitrary URLs, or background surveillance.
Protect credentials.
Nebius and Tavily credentials are stored in Android Keystore-backed encrypted preferences and belong only on the device you control.
Keep an audit trail.
PIRA records allowed and denied action events so decisions and outcomes can be examined.
Stop execution.
An independent emergency stop can halt an active capability flow.
Know the boundary.
PIRA is an experimental prototype with a security-first design, not an independently audited security product.
