What can an investigation agent reach?
Only what your policy grants, and never your estate. Cyberdroid AI Investigation takes every permission from your policy and roster, not from the request or the model.
Controls
Asking is claiming.
Your policy and roster decide who may ask and at what scope, with no model call.
Queries, read-only.
Software turns each question into a read-only query, within set budgets.
One key each.
Each agent has its own revocable key. Revoking one stops one agent.
Recipients checked twice.
Recipients are checked before a reply is written and again after.
Approval, then delivery.
A separate signer approves. You can hold delivery without relying on the model.
Models by role.
Local models are authenticated per key, and routing metadata shows which one answered.
Limits
Estate
No agent can change your estate. Only the suite acts, on your team's rules and approvals.
Policy
Permissions are only as narrow as the policy you write. The system enforces it as written, mistakes included.
Verify
Every grant and refusal is on record.
Outcomes
Each request ends as Task, Pending approval, Refused or Rate-limited, with an operation id and a reason.
Keys
Each key is rotated or revoked on its own, and a call from an unlisted source address is refused.
Standards
Naming a clause is not a claim of certification.
OWASP
LLM06:2025, Excessive Agency, and the 2026 agentic list’s identity and privilege abuse are the risks the first three controls answer.
EU
Article 15 of the EU AI Act asks that high-risk systems resist unauthorized attempts to alter their use. Because asking is only claiming, a request cannot change what an agent may do.
