Our own security protocol,built on a scientific model.
SafeUs does not guess, and it does not hand decisions to an AI. Every action that reaches SafeUs passes one decision point, follows rules you can read and leaves a record you can open. This page explains the model. The exact rules, thresholds and test catalogue stay private, so nobody can tune an attack to slip under them.
Simulation
See how SafeUs checks an AI agent's request
Pick an example. Follow a request from an AI agent through the passport check and see whether SafeUs lets it through, holds it for a person or stops it before it reaches your accounts.
Choose an example
Request path
Hidden instruction in an email
An email hides a line telling the assistant to send every contract to an outside address.
Inbox assistant · send 40 contract files to an outside address
Agent passport check
- Who the agent isPass
- Acting for whomPass
- PurposeCheck
- Allowed accessFail
- Rules and riskFail
Why: sending files outside is not in this agent's passport.
SafeUs decision
Deny
What SafeUs does
- Stops a send that passes through SafeUs before the file leaves
- Quarantines the assistant until you review it
- Shows you the email that carried the instruction
Recorded in the Black Box with the rule that applied.
Across these examples
- Requests checked
- 8
- Held for a person
- 3
- Stopped
- 4
- Inbox assistantFiles & driveDeny
- Scheduling assistantCalendarAllow
- Research agentWebsiteDeny
- Clean-up agentFiles & driveHold
- Support agentChatsHold
The SafeUs protocol
Eight rules the product never breaks.
One decision point
Every request that reaches SafeUs is checked at a single point, in a fixed order, before anything happens. The same request under the same rules always gets the same answer.
Rules decide, not an AI
Clear, fixed rules make the decision. An AI review is used only when something looks suspicious, and it can only make the answer stricter.
Automation tightens, people loosen
Signals, context and automation can only move a decision towards Blocked. Loosening a decision always takes a person and a fresh check.
No record, no action
If the decision cannot be written to the record first, the action does not run.
Zero content
We keep decisions, field names and references. Never the content of mail, files, chats or prompts, and never tokens or passwords.
No proof, no promise
If SafeUs cannot prove it controls something, it says so. The item is marked as your step or outside our reach, never as protected.
High-impact actions wait for you
Sending, deleting, paying or exporting on your behalf always waits for a person to approve it.
Safe by default
The safe setting is the default and is never sold as an extra. Stop is free on every plan.
Decisions
Seven answers. One direction.
Each check gives one of seven answers, from most open to most closed. Any rule, signal or context may push the answer further towards Blocked; none of them can pull it back.
Decision ladder
- Allowed
- Read-only
- Redacted
- Restricted
- Waiting for your approval
- On hold
- Blocked
Proven on every release by thousands of automated tests: no sequence of rules can move a decision back.
Scientific model
Nine questions for every action. No score.
We call it the Action Safety Envelope. Instead of one opaque number, every action is read along nine axes, and each axis is answered by a concrete check rather than an estimate.
Authority
Who is acting, and on whose behalf?
Purpose
Does the request match what was approved?
Scope
Are the action and every field inside what was granted?
Sensitivity
How sensitive is the data involved?
Action power
Can it send, delete, pay or export?
Context
Which device and what kind of network are behind it?
Behaviour
Is this agent acting unlike itself?
Reversibility
Can the result be undone?
Evidence
Can we prove the control really worked?
Hard limits are never weighed
Some situations are not balanced against anything: an unknown or blocked agent, another workspace's data, a forged or expired delegation, a forbidden tool, a missing record. They are blocked or held.
Why there is no score
Risk levels are ranks, not measurements. Multiplying ranks gives numbers that look precise and mean nothing. So SafeUs orders findings by fixed priorities and shows counts and labels, never a number from 0 to 100.
Verified before suspected
A path we have verified always ranks above one we only suspect. You see the three that matter most today first.
Evidence strength
Every item says how sure we are.
SafeUs never calls something protected because it looks fine. Each account, agent and device carries one of four labels, based on evidence we can show.
It passes through a SafeUs control point and the control is proven.
We can see it, but we cannot control it yet.
SafeUs gives you the exact step; you carry it out.
We have no access here, and we say so instead of hiding it.
If the evidence gets weaker, a decision never gets looser. An empty record does not mean you are safe.
Agent power
The more an agent can do, the tighter the controls.
Each agent gets a capability tier from what it declares and what your policy allows. A tier rises automatically; lowering it always needs a person.
- Level 1Reads only, with a single tool.
- Level 2Writes, or uses several tools.
- Level 3Can send, delete, pay, export or reach outside. Every irreversible action waits for you.
- Level 4Runs code, picks up new tools or holds admin rights. Watched by default.
- Level 5Can change or copy itself, or works in health, finance or the public sector. It does not start without human approval.
Testing
We test the model against itself.
Guardian Drill puts your own SafeUs set-up to the test: could this agent, account or path do something it should not? It runs as a safe simulation, sends nothing to your providers, changes nothing and never touches anyone else's systems.
Each drill tells you how many scenarios were tested and how many open paths were found. It never claims you are safe.
Six measures instead of a score
- How often a drill attack gets through
- How often something harmless is held
- Unsafe side effects, with a target of zero
- Time from alert to containment
- How complete the proof is
- How much is still unverified
On a bad day
Contain first, then prove what happened.
Stop
One tap and a confirmation put every agent in quarantine and the workspace in read-only. Free on every plan and never limited.
Stop & Reset
On paid plans: preview, contain, remove SafeUs access, re-check and restore only what is safe. The report has three boxes: what SafeUs did, your step, outside our reach.
Black Box
Before, the moment and after: who asked, who approved, what happened and the proof. Never content.
Safety Check-up
A weekly check that changes nothing: stale access, changed tools and one clear next step per finding.
Our own platform
A security product protects itself first.
Workspace isolation
Every workspace is sealed off from every other. One workspace cannot see another's data, not even by mistake.
Employer privacy boundary
A company workspace holds work scope only. Your personal SafeUs is a separate workspace your employer cannot see.
Hosted in the EU
Database in Ireland, application functions in Dublin.
No passwords to steal
Sign-in uses one-time e-mail codes. High-impact actions ask for a second factor.
Verified messages
Every message from a connected service is verified as genuine and accepted only once, before it can change anything.
Guarded releases
Every change passes automated security checks before it reaches you.
Backups and restore drills
Daily backups. The first restore drill has not been run yet; its result will be published here with the evidence.
What we keep private
The order of checks, thresholds, rule codes and the test catalogue. Publishing them would help someone shape an attack to slip under them. The principles on this page stay fixed; the details behind them keep getting tighter.
Found a flaw? Tell us.
Found a vulnerability? Write to security@safeus.es (copy info@safeus.es). Please do not include access tokens or personal data in the first message.
We aim to acknowledge reports within 72 hours and to keep you informed until the issue is resolved.
Please test only against your own workspace, do not access other people's data and do not degrade the service. Good-faith research that follows these rules will not be pursued.
This page makes no certification claim and shows no score. Where SafeUs has no evidence, it says so.