What stops us from breaking your environment.
Autonomous validation only earns its place if it is safe to point at production. This is the engineering that constrains Radius on a live estate, published so you can assess it before you authorise anything.
Always on
The safety interceptor is not a module.
It cannot be switched off for a faster run, a bigger scope, or a more impressive demo. Every agent tool call passes through all three layers, every time.
The safety interceptor
Three layers of validation, in order, before any action reaches your estate.
01
Hard-coded blocklist
Eighteen non-bypassable patterns that no configuration can switch off — destructive filesystem operations, fork bombs, reverse shells, crypto miners, and anything that would shut a host down.
Cannot be disabled, by us or by you
02
Custom blocklist
Your own patterns, scoped to your tenant and managed by your administrators. Plain text or regular expressions, each rule individually enabled or disabled.
Yours to control
03
Rules of engagement
Per-asset policy: which addresses and ranges are in scope, which ports, which scan types, which exploits are permitted, and the time windows and blackout dates the run must respect.
Agreed before the engagement starts
If any layer blocks a call, the call does not run and the block is recorded with the layer responsible for it.
Every decision is written down
Not only the blocked ones. Each entry records what was attempted, what was decided, why, and — critically — a snapshot of the rules of engagement in force at that moment.
That last field is what makes the trail usable as evidence months later. You are not asked to trust that the policy was correct at the time; the policy is captured alongside the decision it governed.
Recorded per decision
- The action
- Which tool was invoked, and with what input.
- The decision
- Allowed or blocked, and by which layer.
- The reason
- The rule or boundary that produced the outcome.
- The policy in force
- A snapshot of the rules of engagement at that instant.
Measured against published doctrine
Rather than asking you to accept our own account of what responsible testing means, we hold the platform to the boundaries set out in the US GSA’s penetration testing guidance, CIO-IT Security-11-51 Rev 7:
- No action outside the agreed rules-of-engagement boundary.
- No persistence left behind beyond the testing window.
- No new vulnerabilities introduced into the target environment.
- No modification of the target's logs.
To be precise about what this is. CIO-IT Security-11-51 is a recognisable and publicly checkable template for safe testing conduct. It is not a certification, we are not audited against it, and it carries no regulatory authority in Latin America. We cite it because it lets you hold us to somebody else’s definition of responsible behaviour rather than ours.
Questions before you authorise a test?
Send them. We would rather answer a hard security question early than have it surface after the rules of engagement are signed.
Talk to us