It can be — if five things are true. An assistant that can read your email can be fooled by an email, and an assistant that can send can send the wrong thing. The question isn't whether to trust AI. It's whether the product is built so you don't have to.
Five questions to ask before you connect anything
1 · Can it send without you? The safest default is drafts. In Agent07, agents draft and people send. A reply that commits you to a meeting time always waits for your click, and your calendar is re-checked at that click.
2 · Does the rule live in code, or in the prompt? "Always ask before sending" written into an AI's instructions can be forgotten or overridden. In Agent07 the rule is enforced on the server: a reply the agent tries to send on its own is refused, whatever the AI was told.
3 · Can you see what it did? Every agent run is recorded: what it read, which of your rules applied, and who approved the result. Admin and configuration changes go into a tamper-evident audit log.
4 · Does it read only what it needs? When Agent07 finds a time with your teammates, it reads their busy times — start and length — never a title, attendee or description, and only inside your own organisation.
5 · Does it treat other people's words as instructions? Anyone who can email you can write text aimed at your assistant. Agent07 marks emails, transcripts, documents and web pages written by other people as outside material before an AI reads them, and logs anything that looks like an attempt to instruct it.
What failure looks like
| Failure | The guard |
|---|
| An email nobody approved goes out | Drafts by default; sends gated on the server |
| A "confirm first" instruction gets lost mid-task | The rule is in code, not the prompt |
| A meeting is accepted at a time that was just taken | The calendar is checked again at your click |
| An email tells the assistant to forward your files | Outside text is marked as data, and the send still waits for you |
Don't ask an assistant to be careful. Ask it to show you where the care is built in.
How Agent07 handles your data →
Frequently asked questions
- Is it safe to give an AI assistant access to my inbox and calendar?
- It can be, if five things are true: it cannot send without your approval; that rule is enforced in code, not just written in the AI's instructions; you can see a record of what it did; it reads only what it needs (for example, busy times rather than meeting details); and it treats emails written by other people as data, not as instructions. Ask any assistant about all five before connecting it.
- Can an email trick an AI assistant into doing something?
- It is a real risk: anyone who can email you can write text aimed at your assistant. Agent07 marks text written by other people — emails, transcripts, documents, web pages — as outside material before an AI reads it, and logs anything that looks like an instruction. Sending is also gated in code, so a tricked draft still waits for you.
- Does Agent07 send emails without asking?
- Agents draft; they don't send on their own. If you switch on Auto mode — for simple replies, or for paced outbound email — that is your choice, and it runs within limits you set. A reply that commits you to a meeting time always waits for your click.