Security
Security that doesn’t depend on the AI behaving
Rallied’s controls sit outside the AI, in infrastructure it can’t bypass. Asking a model nicely is a prompt. A wall in front of it is a control.
How it’s built
Built as if the AI will get it wrong
That’s the design assumption behind every layer: the model can be confused, manipulated, or plain wrong, and nothing bad should happen when it is. These four decisions are what that looks like.
01
One client’s data can’t end up in another’s ticket
Every client gets its own dedicated agent, on infrastructure dedicated to your MSP. We didn’t point one shared model at all your tenants and hope a system prompt keeps them apart: there is no shared agent to leak through.
02
Nothing runs without your say-so
Anything that changes a client’s environment waits for a human until you decide otherwise, and “reads” like credential or recovery-key lookups are gated too, because read-only is a lie when the read is a password. We picked the slower default on purpose.
03
What you approve is exactly what runs
The agent can’t skip an approval, because approvals aren’t instructions to the AI. They’re a wall in front of it: each step is checked against the plan you signed off, an ambiguous reply classifies to neutral instead of yes, and the agent can’t pick its own approver.
04
You can check everything it did, misses included
Every action, approval, and failure is on the record, per client and per ticket. The failures are surfaced with a reason on purpose, because an audit trail that hides the misses is marketing, and you’d find out eventually anyway.
The rest
And the boring stuff, too
- Password and MFA resets verify the requester with a push to their phone first, so a hacked mailbox can’t reset a password.
- Temporary passwords travel through one-time links that self-destruct. No secret ever sits in a ticket.
- Agents reach out to your tools; nothing on the internet reaches into the agents.
- Sensitive clients can run tighter rules than the rest, and you can take any permission back.
- High-impact changes can require two sign-offs: your technician and the client’s designated approver.
- Inbound webhooks are signature-verified before anything gets processed.
- Dashboard access is role-based and SSO-capable, so a tech can’t widen their own blast radius.
- SOC 2 Type II certified and GDPR ready.
- Your client data doesn’t train shared models.
- Integration credentials are encrypted at rest.
The easiest way to check any of this: Plan Mode runs read-only on your real tickets. Watch it be right before it can touch anything.
Show us the ticket your team is tired of seeing
Bring a shared mailbox request, a new-user setup, or another repeat ticket. We’ll show you how Rallied closes it, from request to done.