New feature · BetaAgent Armor
Guard your agent from its own inbox.
Mail filters ask
Would a human be fooled?
Agent Armor asks
Would a model be steered?
Starts in observe-only. Enable holds when you’re ready.
Inbound screening plane
Safe mail continues. Suspicious instructions stop.
What a person skims past, a model executes.
A normal invoice reply. Switch it to MODEL. Three things are hiding in it.
Read this as
3 hidden →
NS
Priya Raman
11:04
billing@northgate-supply.com
Re: Invoice 4471 (updated remittance)
What Agent Armor sees
Nothing a person would catch.
01
Mail still arrives.
Armor screens after receipt, never in front of your MX record.
2502.0.0 · accepted at the door
02
Armor reads what models read.
Rules catch hidden signals. A model reads the whole message for intent.
homoglyph senderzero-width runimperative to assistantquoted-thread directive
03
Suspicious mail stops before your agent.
When enforcement is enabled, only the webhook waits. Nothing bounces or disappears.
smtp250 unchanged
message.receivedheld for review
Observe first
See what’s already in your mail.
Starts in observe-only. Enable holds when you’re ready.