White paper

The AI prepares. You approve. Pandore applies, and keeps the proof.

The other pages show what Pandore does. This one explains why you can trust it with your infrastructure: the AI prepares changes but never applies them itself, nothing can be altered after you approve, and everything stays provable. It is written for CISOs, architects and operations leads.

October 2026 edition

The AI applies nothing itself
The assistant prepares changes. It has no write access: a separate executor applies them, after your approval.
What you approve is what runs
Once approved, the plan is locked. No one, neither the AI nor a person, can change it before or during execution.
Everything is recorded, nothing is erased
Every request, decision and change goes into a tamper-evident log: a deleted or edited entry shows.

01

Who is allowed to change your infrastructure

Most AI tools for infrastructure let the AI act directly: it gets your credentials and runs whatever it generates, at the moment it generates it. Pandore splits the roles. The assistant prepares a plan. Your rules and an authorized person approve it. Then a separate component, the executor, applies exactly that plan. Only the executor holds the credentials.

Who changes the infrastructure
Typical AI toolThe AI, with your credentials
PandoreA separate executor. The AI never holds the credentials
What gets applied
Typical AI toolWhatever the AI generates when it acts
PandoreExactly the plan you reviewed
Who authorizes
Typical AI toolThe AI, or the person who trusts it
PandoreYour rules, then an authorized person
How the change travels
Typical AI toolStraight onto the servers
PandoreThrough your usual tools, for example a pull request in git
Know-how
Typical AI toolExplained again at every request
PandoreYour procedures, saved and reused
Evidence
Typical AI toolThe chat history
PandoreA tamper-evident log of every action

Propose and decide · read-only

AI assistant

Reads your estate and drafts the plan.

Read-only

Your policies

Allow, bound or block every step.

Before execution

Your approval

An authorized person approves. The plan is then frozen.

Plan locked

Apply · the only write access

Executor

The only one with the key: ephemeral access, scoped to the approved target.

Holds the key

Your infrastructure

The approved change is applied, exactly once.

Applied
Every step is written to the chained log, exportable for audit.

The AI prepares, but it neither decides nor executes. So you can switch AI models, or host one yourself, without touching your rules, your credentials or your history.

02

Built-in safeguards

Every request follows the same loop, from reading your infrastructure to the final report. The safeguards below apply every time. They are not options in a menu: they are part of how Pandore is built.

The AI reads, it does not write

The assistant sees your infrastructure read-only. Write access stays with the executor and is never handed to it.

The approved plan is locked

Once approved, the plan cannot change. A digital fingerprint (hash) guarantees that what is applied is exactly what was reviewed. At the slightest difference, the step stops.

Runs only once

An approved plan runs once. An interrupted step is never restarted automatically.

No improvised commands

The executor does not type free-form command lines. It only makes calls defined in advance, with checked parameters.

Short, narrow access

For each run, the executor gets temporary, single-use access, limited to the approved target.

Read or write, system by system

You choose for each connected system whether it can be changed. A change aimed at a read-only system is refused outright.

Approvals that expire

Your rules decide what can run on its own and what waits for approval, by environment and type of action. A plan not approved in time expires. An irreversible production change can require two approvals.

Secrets never exposed

Passwords and keys are never stored in clear, and they are masked in the log.

03

Pandore plugs into what you already have

No prior migration, no new tool to force on your teams. Pandore connects to what you already run, in six ways. The first five work anywhere, whatever your cloud or data center.

Kubernetes clusters

Any cluster, whoever hosts it.

Servers over SSH

A whole fleet of machines, bastion included. This is often how older systems are reached.

Git repositories

GitHub, GitLab, Gitea. Also how changes travel when your infrastructure is described in git.

APIs

Your internal services or third-party ones, with a rule per endpoint.

SQL databases

Detected automatically, with separate rights to read and to change.

Cloud accounts

Inventory and resource creation after approval, on OVHcloud, Scaleway, AWS, Azure and GCP.

If your infrastructure is managed in git, Pandore goes through git

In that case, Pandore changes nothing directly. It opens a pull request in your repository. Your teams review it under their usual rules, and Pandore is not allowed to merge it itself. Your approval processes stay yours.

04

Your procedures become reusable

Pandore remembers your infrastructure and the way you work, at four levels. When they conflict, a procedure always wins over a note.

Procedures
Your rules and methods, approved by an administrator and versioned. Pandore applies them, it does not just look them up.
Inventory
The up-to-date state of everything you run: clusters, servers, cloud accounts, repositories.
Project notes
What one job teaches the next, such as a workaround found or a mistake not to repeat. An administrator can remove them.
Log
The history of everything requested, decided and applied.

A setup done right once can become a procedure, which Pandore repeats identically for the next site, region or customer. You pay for the design once. And when an expert moves on, their method stays in the company.

05

A log that serves as evidence

Every action goes into a log where each entry is linked to the previous one by a digital fingerprint. So no entry can be deleted or edited without it showing. A regular history says what happened; this one also proves nothing was erased. It exports for your audits.

  1. #1204Request recordedprevious fingerprint0000…0000
  2. #1205Plan approvedprevious fingerprint9f3a…c21e
  3. #1206Change appliedprevious fingerprint4b7d…08afEntry edited afterwards
  4. #1207Result verifiedprevious fingerprinte16c…5d93Broken link: the edit shows
Each entry carries the fingerprint of the one before. Editing or deleting an entry breaks the chain at that point.

What each entry holds

  1. 1The request, and who made it
  2. 2The plan the assistant proposed
  3. 3The rule applied and its decision, refusals included
  4. 4Who approved, and when
  5. 5What was actually applied
  6. 6The verified result

What it brings for compliance

GDPR

Data about your infrastructure stays in the country you choose, and secrets never appear in the log.

DORA and NIS2

For every change to a critical system, you know which rule applied and who approved. You can find everything by date, environment or person.

EU AI Act

When the AI prepares an action, the log links the request, the proposal, the human approval and the result. The record is made when the action happens, not pieced together afterwards.

A log makes audits easier, but it is not compliance on its own, and it only covers operations that went through Pandore. SOC 2 Type 1 certification is planned for 2027: we would rather show you where we stand on a call than announce a certification we do not hold yet.

See Pandore at work, on your own use cases

In a demo, you make a request, Pandore plans it, checks your rules and waits for your go-ahead. On your clouds, your GPUs and your compliance constraints.