Your incidents never leave.Nothing changes without a human.

- inbound ports:0 open
- license:signature verified
- ledger seal:intact
Letting software near production should worry you. So here is everything your security team will ask, in the order they’ll ask it.
First, can anyone get in? No.
A Helm chart or a single host, inside your network. Nothing listens from outside.
3AM runs inside your network and opens no inbound ports. Probe every one of them from outside, and none answer. It only ever calls out, and you decide how far.It runs inside your network, with no inbound ports. Probe them all: none answer.
Read the architectureNext, what does it see? Never your secrets.
Logs are full of things that shouldn’t travel. So passwords, tokens, keys and card numbers are redacted on the way in, before anything is written. Prompts are kept redacted too, with a hash for proof.Passwords, tokens, keys and card numbers are redacted on the way in, before anything is written.
- No incident data sent to us
- Models on your GPU, or an endpoint you choose
- Card numbers caught by checksum
- db.connect user=svc_3am password=hunter2-prod!
- slack.post token=xoxb-2291-8812-Kq3vZ channel=#approvals
- payment.retry card=4111 1111 1111 1111 amount=420.00
- git.fetch Authorization: Bearer ghp_9fK2mQx7Lw0
- ssh.exec key=-----BEGIN OPENSSH PRIVATE KEY-----
Then, who can change anything? Two people, always.
Seeing is one thing. Changing is another. Every write takes two people, each with their own hardware key. Code review can't grant that, and neither can an admin. Only their signatures can.Every write takes two people, each with their own key. Code review and admins can't grant that.
- Two people, each with a hardware key
- Code review and admins can’t grant it
- One switch stops everything
- core-banking-dbShadowL1
- payments-edgeShadowL1
- messagingShadowL1
- core-bankingShadowL1
And whatever happened, you can prove it.
Months later, an auditor asks what happened at 03:14. Every signal, check, approval and action is hash-chained, and your own key co-signs each seal, so nobody can quietly rewrite it. Not even us. Change one byte and the verifier shows where.Every action is hash-chained and co-signed by your key, so nobody can rewrite it. Not even us. Change one byte and the verifier shows where.
- Every signal, check, approval and action, chained
- Your own key co-signs each seal
- One changed byte shows where
Checking
Each line holds the hash of the one before, so the verifier walks them in order.
0 of 6 lines verified
Click any line to change one byte in it. Illustrative data.
One last worry: what if the AI gets tricked?
How it worksA log line, a ticket or a row can hide an instruction, and it might change what the AI suggests. That’s survivable, because no suggestion can pass a gate. Only a live check can authorize a fix.A hidden instruction might change what the AI suggests. Only a live check can authorize a fix.
Can suggest
Read for ideas. Never enough to act on.
Runbooks and wiki pages · Past postmortems
Can authorize
Only a live check can authorize a fix.
Code and schema · Alert rules and live metrics
Is only data
Treated as data, never as instructions. Your approval policy still applies.
Instructions hidden in text
- Builds you can verify yourself
- Signed, distroless images
- SBOM with every release
- Signed license, verified offline
- Source-free compiled build
- Read-only diagnostic credentials
- No shell access for the agent
Now assume all of that fails. Including us.
Keys get stolen and vendors get hacked, and we’re a vendor. So take each place an attacker could land, and ask what they’d get. None of them is enough to change your data or loosen the rules.Take each place an attacker could land. None is enough to change your data or loosen the rules.
1 of 5
stopped
Your data and your rules stay as they were.
So locking takes a second. Unlocking takes days.
An attacker's first move is to loosen the rules, so that's the move that waits. Tightening them never does: any one person can do it, the moment something looks wrong.An attacker's first move is to loosen the rules, so that's the move that waits.
Request: Let it do more
Several signatures, then a wait. Never the person who benefits.
72:00:00left to wait
Making it safer
Any one person, immediately.
Making it riskier
Several signatures, then a wait. Never the person who benefits.
While it waits, everyone is alerted, and any one of them can cancel it. Your clock times it, not ours.While it waits, everyone is alerted, and any one of them can cancel.
Suggested defaults. You set your own, above fixed floors.
Your CISO
Before the pilot
Can anyone outside reach it?
No. It opens no inbound ports. It calls out only to the tools you connect, and to a signed license check.
- 0inbound ports
- 2people for every write
- 1switch stops everything
- 1 bytechanged, and the ledger shows where
What your security team asks next
You do. At install, named people from your bank each enroll a hardware key and sign the first policy: who may approve, which tables are the books, and how many signatures each change takes. 3AM staff are never among them.
Yes, on every license tier. It signs in through OIDC, including Entra ID, ADFS, Okta, Ping and Keycloak, or directly against LDAP or Active Directory over TLS. Your provider's MFA and access rules apply, and directory groups decide each person's role.
There are four roles: owner, admin, responder and viewer. A viewer can read incidents and the audit log, and can see connections but never their secrets. Someone in no mapped directory group gets no access.
The license service sees fingerprints, public keys and licenses. It never sees a connection, a secret or a person. Air-gapped, it sees nothing at all.
Text inside anything it reads is treated as data, never as instructions. A fix still needs a cause confirmed by a live check, and your approval policy still applies.
Yes. Every model call goes on the ledger: the model, the redacted prompt with its SHA-256, the output, tokens and latency. Whoever holds the original prompt can prove which one it was.
It runs as non-root with a read-only root filesystem, all Linux capabilities dropped, seccomp RuntimeDefault and no service-account token. An in-cluster model is reachable only from the core, by NetworkPolicy.
A license can name your DR install, or carry grace hours for failover. On Enterprise and Strategic, Keychain sets up every install once, and that setup moves only between your own installs.
Bring your security questionnaire
- Stays in your network
- Humans hold the keys
- Tamper-evident
- Shadow first