SECURITY
Control is the product.
An AI that only chats needs little trust. An AI that acts on your behalf needs verifiable control. So control is not a settings page in DRENTIR — it is the architecture: permissions define what each agent may do, verification checks every result, and the audit trail records it all.
Design approach — illustrated, not production software. No security certifications are claimed.
PERMISSION LEVELS
You decide how far it may go.
Every agent operates at a permission level you set. Higher reach is not assumed — it is granted.
L1
READ
Observe only. At this level DRENTIR looks at your data — it does not touch, change, or move it.
e.g. Reads incoming emails and summarizes them for you.
L2
PREPARE
Do the work, but stop before anything leaves the system. Drafts, plans, proposals — all held for your review.
e.g. Drafts replies to the emails it read.
L3
ASK BEFORE ACTION
Everything is ready. Nothing happens until you approve — explicitly, every time.
e.g. Asks for your OK before sending a reply.
L4
AUTONOMOUS WITHIN POLICY
Act freely inside boundaries you defined in advance. Policies are written by you, reviewable, and revocable.
e.g. Files routine reports automatically — the way you told it to.
IN PRACTICE
What each level may actually do.
Irreversible actions require your approval by design. Routine work flows — inside policies you wrote.
Defaults are conservative. You raise a level deliberately — never the other way around.
VERIFICATION
Nothing acts unchecked.
DRENTIR checks itself before anything becomes final — not after.
- Every plan is reviewed against your permission boundaries before the first step runs.
- Every result is checked against the original instruction before it becomes final.
- Anything irreversible pauses for your explicit approval — every time.
- When DRENTIR is unsure, it asks instead of guessing. Uncertainty is reported, not hidden.
AUDIT TRAIL
Every action, on record.
What happened, when, and on whose instruction — reviewable any time. Your log is yours.
-
Read 3 new emails
L1 · READ -
Drafted a reply
L2 · PREPARE — held for review -
Waiting: your approval to send
L3 · ASK BEFORE ACTION -
Approved by you — sent
Recorded with your instruction attached
Illustration of the log format.
SECURITY STATUS
What is built, what is invariant, what is next.
Security status: architecture implemented in parts; validation ongoing. We claim no certifications and no absolute guarantees.
A · Implemented now
- User authentication foundation
- Permission scopes
- Workspace separation foundation
- Audit events
- Local model isolation where applicable
B · Architectural invariants
- The LLM is not a security boundary
- Tenant identity is injected below the LLM layer
- Agents do not receive master credentials
- Privileged actions require policy evaluation
- Cross-workspace access is denied by infrastructure
C · Planned validation
- Automated cross-tenant security tests
- Prompt injection testing
- Malicious tool / MCP testing
- SSRF testing
- Credential exfiltration testing
- External security review before public launch
Found a security issue? Email us — it goes to the front of the queue.
QUESTIONS
Asked about control.
Not by default. Sending anything — an email, a message, a file — requires the ASK BEFORE ACTION level, and that level pauses for your explicit approval every time. Fully automatic sending is only possible inside narrow policies you write yourself, and you can revoke them at any moment.
What you delegate, the permissions you grant, and the audit trail of what was done. Data exists to do your tasks and to let you review them — nothing more. You can see what is stored and ask for it to be removed.
Each agent carries its own permission level and its own data scope. An agent cannot reach data it was never granted, borrow another agent's access, or escalate its own level. Sharing between agents requires your explicit instruction — and it lands in the audit trail.
Email security@art-webstore.com with “Security report” in the subject — it goes to the front of the queue. We take every report seriously and answer every one. Contact us