Content arrives
A PDF, a note, a web page, a message from a connected account. Anything QP might read.
Fenced as data
Passages quoted to the model are normalized, markup-escaped, and wrapped in a block labeled as material to read, not orders to follow.
QP answers, citing the passage.
Any outward action waits for your yes.
What protects you today
Four layers. Each one checkable.
Any single filter can be fooled. So AskQP stacks simple defenses you can inspect, and keeps the last word with you.
Fenced passages
Text from your files reaches the model inside a labeled block, with markup escaped, so it reads as material rather than as instructions.
Pattern scan on connected content
Content from connected services is checked for known injection phrases, hidden text, and invisible characters before QP reads it.
Model policy per Space
Each Space decides which models may see its content. The router enforces the choice and fails closed.
Direct, guarded web reads
Web pages are fetched straight from the site. Addresses that point back into your own network are refused.
How it works
Four checks. On connected content.
Normalize
Unicode is put into one canonical form, so look-alike and invisible characters show their true shape.
Detect
Known injection phrases, hidden-text tricks in HTML, and invisible characters are matched and flagged.
Score
The signals combine into one risk level for the content.
Fence
The content is escaped and wrapped in a labeled block before the model reads it.
The last word is yours.
Screening narrows the risk. Your approval closes it. Sending mail, creating tasks, writing to a panel, and running a multi-step job all wait for your yes. Every governed change returns a receipt.
Read
See exactly what QP prepared
Approve
It goes out with your yes
Decline
It stays where it is