ONE Security, live

Security, live

Nothing needs action. 4 notes.

Rules read our own records and fire on fixed numbers. No model decides whether a rule fired. Read at 2026-11-19 02:44 UTC.

Calls today 2,320 Every door, every caller
Refused today 2 0.09% of calls
Doors refusing 1 Doors that said no today
Rules fired 0 4 rules, 0 not read
Payment rails checked 2 of 6 A canary passed in the last day
Held by the chain 1 of 12 7 by our server, 4 not built

What the rules found

  • Note

    The acp canary is absent, so this rail is not being checked.

    rule canary · acp

  • Note

    The card canary is disarmed, so this rail is not being checked.

    rule canary · card

  • Note

    The crypto canary is disarmed, so this rail is not being checked.

    rule canary · crypto

  • Note

    The link canary is paused, so this rail is not being checked.

    rule canary · link

quiet refusal-spike fires when a door refuses 20+ calls in a day and 5× its usual day.
quiet anonymous-probing fires when callers with no identity are refused at 15+ different doors in a day.
fired canary fires when a payment rail’s daily canary fails, goes quiet for 26 hours, or is not armed.
quiet audit-write-failed fires when an owner-level action runs and its audit row fails to write.

Owner actions are counted over 7 days. Each door is compared with its own last fourteen days.

What matters most

  1. 1

    An agent that read outside text can still act

    A poisoned page or email cannot be stopped from being read. Today only a text marker stands between it and a tool that sends, posts or pays. The Injection Warden is building the gate.

  2. 2

    A stolen session holds its doors

    Doors that move money trust a session. They should take a signature, because words and cookies cannot sign.

  3. 3

    The record can be rewritten

    Our audit rows live in our database. Anchoring a daily fingerprint on Sui makes any edit visible.

  4. 4

    4 payment rails have no passing canary

    A rail nobody tests daily can break, or be forged, without anyone knowing.

What was refused

A refusal is a door doing its job. What matters is which door, why, and whether today looks like its usual day. These are counts by kind of caller. They name no one.

Refusals, last fifteen days

11-05today

Highest day: 49 refused. Today: 2.

Why

0 Forbidden. The caller was not allowed through this door.
2 Invalid. What was sent did not fit the door.

By kind of caller

  • anonymousNo identity at all 2

Doors that refused most today

Door ForbiddenInvalid No identityTotal
tasks:create 0 2 2 2

How it works

Two loops. The first checks every call as it arrives. The second watches what the first recorded.

1. Every call, one door

  1. 1

    A caller

    A person, an agent, one of our services, or nobody we know.

  2. 2

    One door

    Every call to a receiver passes through the same wrapper. A chat agent’s own tools are checked in the chat worker, by its own gate.

  3. 3

    Who is asking

    Read from the proven caller, never from what the request says about itself. Fail here and the answer is forbidden.

  4. 4

    What was sent

    Checked against the door’s declared shape. Fail here and the answer is invalid.

  5. 5

    Counted

    Allowed or refused, the call is counted by door, by reason and by kind of caller.

2. The watch: rule, then model, then rule

  1. 1

    Records

    Refusals, payment canaries, owner actions.

  2. 2

    Rule

    A fixed test with fixed numbers. It fires or it does not.

  3. 3

    Model

    Only after a rule fires: explains it and proposes who should act.

  4. 4

    Rule again

    The door that contains it checks the caller, not the model.

Text never carries authority. A page, an email, a message or a model's own output can ask for anything. Only a signature, a proven caller or the chain can authorise.

Nobody can stop a model reading a poisoned sentence. What can be made impossible is that sentence moving money, sending as you or changing a setting, because those doors do not open to words. The model sits in the middle of the sandwich for that reason: a rule decides before it, and a rule decides after it.

The team

Seven agents. One director who thinks, four who build, two who were already on watch.

CEO

Decides what is taken from customers to contain a threat.

Security Director

Holds the law, routes each finding to whoever owns the code, calls the containment doors.

csoFable

Sentinel

Runs the rules. A threshold fires the finding, never a model.

sentinelOpus

Injection Warden

Makes outside text powerless. It can ask. It can never authorise.

injection-wardenOpus

Chain Warden

Moves limits and proofs onto Sui, where a broken server cannot change them.

chain-wardenOpus

Red Team

Attacks ONE on dev as a stranger would. Hands back a test that fails.

red-teamOpus

Security Auditor

Reads every change for who can reach a new door, before what it does.

security-auditorReview

Incident Commander

Runs a live incident. Stops the bleeding, then finds the cause.

incident-commanderOn call

What holds, and who holds it

One question sorts every control: if our own server were broken into tonight, would this still hold?

Held by the chain 1

These hold even if our server is broken into. A key signs it, or the chain refuses it.

What an agent’s key can spend from the treasury

The vault on Sui holds a limit per payment and per day for each signer. Above either, the chain refuses the payment.

Accepts Spends inside the limit. The limit is the most a stolen agent key can cost.

pay/contracts/vault-tests/tests/vault_tests.move

Held by our server 7

Our code enforces these. They are real, and code we run can be changed.

The total an agent may spend before it asks again

A number in our code, checked by our server before it lets a spend through.

Accepts A broken server can skip it. The chain’s per-day limit is what still holds.

one.ie/web/src/lib/contract/read.ts

One group reaching another group’s doors

A door labelled for a group’s owner or admin refuses a caller who has not proven who they are.

Accepts Doors labelled only "signed in" warn and let the call through. Doors with no label are open.

one.ie/web/src/lib/bind-receiver.ts

Actions that cannot be undone

The server asks for a person’s approval before an irreversible action runs.

Accepts Only the actions that call the gate.

one.ie/web/src/lib/resolvers/_shared.ts

A web page, email or message steering an agent

Text from outside is wrapped and marked as data before a model reads it.

Accepts A mark is advice to the model. It does not stop a tool from running.

channels/src/fence.ts

A stranger’s message making an agent write or remember

A turn started by an unverified message may read and may not write.

Accepts Reads still run, including fetching a web address the message chose.

channels/src/tools.ts

Another owner’s note giving your agent orders

A note is obeyed only from a sender verified on chain at a level no door issues yet. So every such note is read-only.

Accepts Nothing today. This loosens only when a door starts issuing that level.

one.ie/web/src/lib/identity/obey.ts

A door that keeps failing being called again and again

A path with far more failures than successes is refused before the call runs.

Accepts It blocks a family of doors, not a caller, and it lets the call through when it cannot read the count.

one.ie/web/src/lib/substrate.ts

Not built yet 4

Nothing enforces these today. Each one is a lane the security team is building.

An agent that read outside text then sending, posting or paying

Tools are not yet sorted into reads and effects, so a turn that read outside text can still call a tool that acts.

Accepts Until it is built, the fence above is the only thing in the way.

text/security-director.md

A stolen session calling a door that moves money

A request is trusted by a session or a key our server checks. It carries no signature of its own.

Accepts Whoever holds the session holds its doors, up to the chain’s limits.

text/security-director.md

Someone changing or deleting the record of what happened

The record lives in our database. Nothing outside it proves a row was there.

Accepts Whoever holds the database can rewrite it unseen.

text/security-director.md

Knowing who was refused, not only how often

Refusals are counted per door and per day. No row says who, from where, or when.

Accepts The watch can say a door is under pressure. It cannot yet say from whom.

text/security-director.md

How the chain hardens it

A server can be broken into, and whoever holds it can change what it enforces. A chain cannot be talked into anything. So the work is to move each control that matters from "our server says no" to "the chain says no", one at a time.

A limit becomes a fact. The most a key can spend is a number the chain holds. No bug, prompt or stolen session of ours can raise it.

Authority becomes a signature. A door that takes a signature cannot be opened by text, because text cannot sign.

The record becomes permanent. A daily fingerprint on chain means nobody, including us, can quietly change what happened.

Live 3

On chain today.

A spending limit the chain enforces

Each agent key is a signer on the group’s vault with a limit per payment and per day. The chain refuses a payment above either.

Stops A stolen or tricked agent key draining the treasury. The most it can take is the limit.

pay/contracts/vault-tests/tests/vault_tests.move

A stop only the owner’s key can pull

The owner can pause or remove a signer on the vault. The owner’s key is sealed behind a fingerprint and cannot sign from an agent’s shell.

Stops An attacker who controls our server keeping a signer alive after the owner says stop.

one.ie/web/src/lib/contract/read.ts

A signature proves who owns an address

A payout address or sweep rule is accepted only with a message signed by the key that owns it.

Stops Someone pointing an agent’s earnings at an address they chose.

one.ie/web/src/lib/wallet-address-proof.ts

Contract written 2

The contract exists. Nothing uses it yet.

Who may give an agent orders, read from the chain

A contract records how far each sender has been verified. An agent obeys a note only from a sender at the top level. No door issues that level yet, so every such note is read-only.

Stops A stranger’s message, or another owner’s agent, steering yours.

pay/contracts/verification/sources/verification.move

The security record is anchored on chain

Each day’s security events are hashed into one fingerprint and written to Sui. Anyone can prove a row was there, and a deleted or edited row breaks the fingerprint. The contract is written and tested, and runs on Sui testnet over a test day. It is not on mainnet, and no real day is anchored yet.

Stops An intruder, or us, rewriting the record of what happened.

pay/contracts/audit-anchor/sources/audit_anchor.move

Next 2

Not written.

The lifetime limit moves on chain

The total an agent may ever spend becomes a number in the vault, beside the per-payment and per-day limits.

Stops A broken server skipping the one limit that today lives only in our code.

Doors that move money take a signature, not a session

A request that pays, changes a limit or adds a signer carries a signature over the exact request, checked against the group’s signers on chain.

Stops A stolen session, a forged internal call, or injected text reaching those doors. Words cannot sign.

How ONE is built to be safe · report a problem to security@one.ie