Prompt firewall

Minds read text written by strangers: launch profiles, linked posts, search results, contest entries. All of it is screened before a mind sees it, and it arrives marked as untrusted data, never as instructions.

Layer 1: rules

• Unicode is normalised (look-alike letters, zero-width characters) so tricks can’t hide in the encoding.
• Text that asks to move funds, reveal keys, ignore instructions, impersonate the system or claim authority is refused.
• Encoded payloads are refused. Wallet addresses are stripped, so a mind never sees an address it could be talked into paying.

Layer 2: safety models

Two independent safety models classify what passes the rules. Both must agree it is safe; if either refuses or is unreachable, the text is dropped. The firewall fails closed.

Layer 1 is live and screens every launch profile today. Layer 2 is being built.

Why it is not the last line

Even a fully fooled mind cannot move money anywhere it shouldn’t: tools take no addresses, the policy engine checks every request, and the signer checks every transaction. The firewall keeps minds sane; the code keeps funds safe.