Read the transcript
Every agent you put to work is a new way in. So if one were compromised, what could it actually reach?
Blast radius answers that, for every agent. It maps each path: what the agent has used, what it could use on its next turn, and what’s one human step away.
Every path is graded by severity, so the critical ones stand out. This one can run code.
Preview the fix, and watch its reach pull in. Then deny it, and it’s out of reach from the very next call.
It reads your own records and never calls out, so it works even air-gapped.
Know your exposure before anything goes wrong, and shrink it in seconds. That’s why your security team can say yes.
That’s governed autonomy. That’s Decision Terrain.
What happens in the demo?
The run follows one agent from its full map of reachable paths to a smaller one, with a single change made by a person.
- Open one agent’s blast radius. The view covers a single agent, Ledger, and maps every path it has to people, data and credentials, model providers, integrations, and communication tools.
- Read the three rings. Observed paths have recorded use in retained runs or data flow. Reachable-now paths are usable on the agent’s next turn with no change by a person. Dormant paths are configured but one human step from use.
- Find the critical paths. Each path is graded by severity. python_exec is marked Critical: if the agent were compromised, it could run code in the agent’s sandbox and reach any host the sandbox network allows.
- Preview the fix. Previewing a denial shows how the agent’s reach would change before any setting is saved.
- Deny it for this agent. The tool is denied for Ledger. Its state changes to Denied for this agent, and the reachable-now count drops.
Why look at reach before something goes wrong?
Security reviews usually ask what an agent is allowed to do. Blast radius asks the harder question: if the agent were manipulated or compromised, what could it get to, and how directly? Seeing that per agent, with each path graded, lets a security team remove the dangerous paths before an agent goes into production instead of after an incident.
The view is built from the deployment’s saved configuration and retained audit records. It does not contact a model provider or outside service to draw the map.
What the demo shows, and what it does not
- The organization, agent, and data are synthetic. No real customer or program is shown.
- The denial was applied for real on camera and reversed after recording. The change history keeps both steps.
- Map labels for a model gateway, an email address, and a local address are blurred.
- Severity reflects what a path would allow if the agent were compromised. It is not a finding that the agent misbehaved.
- Paths outside what the deployment records, such as what a tool does after it is called, are marked as not observable rather than assumed safe.