Security work you shouldn't have to think about
Connection credentials are encrypted where they're stored, assistants browsing the web can't reach anywhere private, and a few doors that were merely closed are now locked.
· Security
Most of what shipped this week is invisible if it’s working. Worth saying out loud anyway, because “trust us” is not an argument.
Your connection credentials are encrypted where they’re stored
When you connect an account, Ward holds a scoped token — never your password. Those tokens are now encrypted before they’re stored, so a leaked copy of our storage yields unreadable data rather than working credentials.
The assistants themselves never see the raw credential. They ask Ward to act on a connection; Ward holds the key.
An assistant on the open web can’t reach anywhere private
Assistants that fetch web pages are now fenced in: they can reach the public internet and nothing else. They can’t be talked into requesting something on our internal network, and they carry no credentials into a page they fetch.
This matters because a page an assistant reads is untrusted input. If a page tries to instruct the assistant to go somewhere it shouldn’t, there’s nowhere for it to go.
Doors that were closed are now locked
A handful of entry points that were previously validated now fail closed — if something can’t be verified, it’s refused rather than accepted with a warning. Errors also no longer echo internal detail back in a response.
None of this changes anything you do. It changes what happens on the bad day.
You can read the full picture on the security overview, including what isn’t in place yet.