AgentiCCells
Choose a designDesign
Idioma: Español (México)

Esta página todavía no está traducida. Estás leyendo el original en inglés.

Trust and security

This page is for the person who has to decide whether a cell is safe to put near company systems. It says what we can show, what we cannot, and where to check. Each claim links to a public file in our repository. If a claim has no link, treat it as unproven. Nothing here has been checked by a third party.

See the website cell

What our agents can reach, and cannot

What a cell's agents can reachFive reaches, from narrow to wide: self, project, host, tenant, world. Each box links to the policy file that lists what was measured at that reach.Selfits own sessionProjectits worktree of this repositoryHostthe one shared containerTenantaccounts its keys and connectors reachWorldany host on the internet, and the live site
What our own agents could reach on 15 November 2026, narrow at the top, wide at the bottom. It is the measurement of this team in one shared container, from data/abp.json; the text below says where it has changed. Today the limits are mostly rules, not walls.

We measured it, and the measurement is unflattering. On 15 November 2026 the agents on this team were granted 20 of the 23 capabilities in the vocabulary we use, and the widest reach was "world" (any host on the internet). Each role was asked to use only 8 to 11 of them. Every "may not" line in a role's brief is a rule written in prose, not a boundary a machine enforces. The figures, role by role, are in data/abp.json and board/abp/README.md, derived by a script (derive.mjs) that fails the build if someone edits the result by hand.

What has changed since: each role other than the conductor now runs as a fresh session in its own container, so one agent cannot read another's keys (the "one container per run" section). What has not changed: the cloud credentials in an agent's environment, and the connectors that could send mail. The same README lists these as the three biggest gaps, each with a proposed narrowing that needs Dinis Cruz's decision. None is closed.

In plain words: today the limits on our agents are mostly rules they follow, not walls. We say so on the questions page, and we will not describe a rule as a guarantee.

Where your data goes

The rule is that we keep no state for customer work. It runs in your environment, and customer data does not come into our accounts. The rule is written for every agent in CLAUDE.md ("Never take customer data into our accounts. We keep no state: customer work runs in the customer's environment.").

Limits: this is a rule, and the measurement above says rules are not yet walls. The one flow where work runs on our side, the website cell, is public by design and has not been run for a customer (FAQ). We have no statement yet on audit-log export or retention for a customer's cell; we have not written one, and will not claim one before we have.

How a release is gated

Every push to the dev branch publishes the site. Before it is published, the workflow runs these, and a failure stops the release:

Our own history scan of 15 November 2026 found no secret shapes in any commit, file name or commit message (board/security/2026-11-15-history-scan.md).

Limit: the same push that is checked can change the workflow that checks it. A branch rule that makes the check a boundary is proposed, not set (board/abp/README.md, gap 3).

Keys and signing

Each agent makes its own key pair when its session starts, with team/keygen.sh. The private half stays in that session and is meant to end with it. Only the public half is committed, in team/keys.json, and the leak scan refuses any private part. Our check on 15 November 2026 found the key folders locked down and the private keys stored encrypted (board/security/2026-11-15-key-hygiene.md).

Limit: the commits on dev are signed by one key in the container's git configuration, and the Agent: trailer is typed by the agent. So the trailer tells you who claims the work, and does not prove it. Signing each agent's commits with its own key is proposed (board/abp/README.md, gap 2).

How to stop everything

A file named board/HOLD.md. While it exists, every agent must do nothing but read (CLAUDE.md: "Never remove board/HOLD.md, and never work while it exists."; the scheduled-run rule is in board/briefs/SCHEDULED.md). Any person with push access can create it. The security role may create it in an emergency (security brief). Limit: it is a file the agents are told to obey. It has not been tested against an agent that ignores it.

What agents never do

Send email, post, publish to social media, contact a person or company, move money, sign up for paid services or accept terms. The rule is in CLAUDE.md: a person sends, the agent drafts. This is a rule (see the first section), not yet a boundary.

Limits we state in public

Two of Dinis Cruz's newsroom articles of 17 November 2026 draw lines we work to. We quote them, and we say which parts are his position and not a product we sell.

Money stays in plain code. In The payment desk he writes: "For mission-critical things like this, you want code that is reliable, auditable, that we can check and verify. You should not be using AI for that." On this team the rule is stricter still: no agent moves money, and no agent has a wallet or a payment power of any kind. The article also says that giving agents wallets, with a small budget, is "my position, not a product we run". We do not offer it.

Keys, regions and erasure, in plain words. Patient vaults, answered answers seven questions from a health team about the vaults we build on. Three of its answers matter to any security reviewer:

What we do not claim: we do not sell patient vaults, a doctor console or an erasure service, and neither article says we do. The article says there is no live date for the doctor console.

What a cell asks of you (proposal)

This section is a proposal. Where nothing is written down yet, it says so.

What we cannot yet show

If you are a security reviewer, tell us which of these you need first. We would rather be asked than guess.