Connecting an AI agent to a security platform is easy. Deciding what it's allowed to do once it's connected is the hard part.
A read is cheap. "List the critical findings on this project" hurts nobody. A write is different. Closing a finding, changing a policy, opening a pull request. You want those, but not from every agent, for every person, in every team.
So Safeguard now checks every MCP tool call against four layers before it runs.
The four layers
- Tenant. What your whole account allows.
- Organization. What a business unit or team allows.
- Role. What a person's role allows.
- Agent. What a specific agent is allowed, separate from the person who set it up.
A call has to pass all four. If it doesn't, the refusal tells you which layer blocked it. No more guessing whether it was a role setting or an org policy.
Reads on, writes off
New tenants now start with read tools on and write tools off. You turn writes on deliberately, where you want them. Existing tenants keep the write tools they already had, so nothing breaks.
Changes you can take back
Widening access is the risky direction, so it asks for confirmation first. And any widening change can be undone in one step.
Every change goes into an access history. You can filter it, check effective access for a given person or agent, and export it to CSV for an auditor.
Why this matters now
Agents are starting to act, not just answer. The question security teams will get asked isn't "do you use AI?" It's "what can your agents change, and who approved that?"
With layered MCP access you can answer both, with an export to prove it.
Try it yourself
Click through it at your own pace, or open it full screen.
List every agent you've connected to a security tool. For each one, write down the single most damaging write action it could take today. If you can't answer for any of them, that agent should start read-only.
Quick check
An agent's write call is refused. How do you find out why?
The refusal names the layer that blocked it: tenant, organization, role or agent.
What do new tenants start with?
Read tools on and write tools off. Existing tenants keep the write tools they already had.
Your auditor asks who widened an agent's access last month. Where's the answer?
In the MCP access history, which you can filter and export to CSV.
