A copilot that proposes; you decide.

HodeiAI queries your organisation's data with tools, explains why a vendor scores the way it does and proposes the next action. It does not act on its own: a person with a write role approves every execution.

Request a demo Read the HodeiAI documentation

Pseudonymises the personal data it recognises before sending
It never acts without a person's explicit confirmation

HodeiAIVendor A

Sample data

Question

Why does Vendor A score so low, and what should I do this week?

Queries to your data

Vendor score and components
Open findings and CVEs in KEV (already exploited)
Blast radius at 3 hops

Answer

It drops because of an unpatched CVE from the KEV list in a component whose impact reaches 6 vendors at 3 hops. I suggest logging the risk so it has an owner and a date.

Proposal

Awaiting approval

Create an entry in the risk register

Approver: User with a write role
Approval: Single use · 15 min
Bound to: Hash 7c1e…a90b of this proposal

Proposal, approval and execution go into an action ledger that only ever adds entries, each linked to the previous one.

Sample data: the vendor, the findings and the proposal are fictional.

What it does with your data

It works with tools that query your organisation at the moment you ask, not with an indexed copy of documents. It is not RAG.

Explains a score
Why a vendor scores the way it does, citing the concrete findings.
Summarises a framework
The status of a compliance framework and which control to tackle first.
Drafts a communication
A draft to a vendor about a finding, which you review before sending it.

Virtual CISO reports are not written by HodeiAI: they are deterministic. Only the advisory panel uses AI.

What goes to the model, and what does not

The path of a question and where each safeguard applies.

HodeiShield
Your organisation
The question and your organisation's data, under its usual isolation.
Inside HodeiShield
Pseudonymisation
Replaces the personal data it recognises. The mapping lives in memory only.
Sub-processor
Mistral AI
Reasons over the pseudonymised text.
Inside HodeiShield
Your answer
The pseudonymisation is reversed inside HodeiShield before you read it.

No user or organisation identifier is sent. We do not claim that no personal data leaves: the data HodeiShield recognises is pseudonymised.

HodeiAI proposes; a person approves; it is recorded

HodeiAI is the proposer in ACT, the action phase of the cycle. This is what it does and what it does not.

What ACT does today

It proposes, and executing requires the explicit approval of a user with a write role.
Each approval is single-use, valid for 15 minutes and bound to the hash of the exact proposal.
Today the only executable action is creating an entry in the risk register.
It records every step in an append-only action ledger, hash-chained and batch-signed with ML-DSA-65, verifiable with the public key.

What it does not do

Execute on its own: autonomous execution is blocked in the database, the API and the executor.
Isolate devices, block accounts, patch, revoke access, open tickets or contact vendors.
Bypass the kill switch: it stops the agent and fails closed.
Prove the decision was the right one: the ledger proves what was proposed, who decided and in what order.

Every claim, with its proof and its scope

What HodeiShield says about HodeiAI and ACT, and where you can check it.

HodeiAI uses Mistral AI models.
Docs · HodeiAI (opens in a new tab)
Scope. Mistral AI is listed as a subprocessor; the detail is on the subprocessors page.
Before anything goes to the model, HodeiShield pseudonymises the personal data it recognises: emails, phone numbers, national IDs, IBANs and the names of people in your organisation.
Docs · Subprocessors (opens in a new tab)
Scope. We do not claim that no personal data leaves: it depends on what it recognises. The risk analysis of supplier scans does not pseudonymise: it sends the supplier's domain and the findings.
HodeiAI works under the same per-organisation isolation as the rest of the application: it does not see or mix another organisation's data.
Docs · HodeiAI (opens in a new tab)
Scope. It queries your data with tools; it does not index it into a document store.
Conversations are kept for 90 days at most and deleted automatically. An administrator can switch HodeiAI off and, while it is off, no call is made to the model.
Docs · HodeiAI (opens in a new tab)
Scope. Without HodeiAI there are no ACT proposals.
HodeiAI proposes. Executing requires the explicit approval of a user with a write role: single-use, valid for 15 minutes and bound to the hash of the exact proposal.
In the demo
Scope. Today the only executable action is creating an entry in the risk register.
The default level is «suggest» and an administrator turns on approval to execute. Autonomous execution is blocked in the database, the API and the executor, and a per-organisation kill switch stops the agent and fails closed.
In the demo
Scope. ACT does not patch, isolate devices, revoke access or contact vendors.
Every proposal, approval, rejection and execution is recorded, with who did it, in an append-only action ledger, hash-chained and batch-signed with ML-DSA-65, verifiable with the public key.
In the demo
Scope. It proves what was proposed, who decided and in what order; not that the decision was the right one.

Keep going

HodeiAI documentation Subprocessors

HodeiAI: straight questions

The ones security leads ask before letting an AI touch their data.

Does it act on its own?

No. HodeiAI proposes and a person with a write role approves every execution, with a single-use approval, valid for 15 minutes and bound to the proposal's hash. Today the only executable action is creating an entry in the risk register. Autonomous execution is blocked and the kill switch fails closed.

Which model does it use?

Mistral AI models; Mistral AI is a sub-processor of Hodeitek.

Is personal data sent to the model?

Before sending, the personal data HodeiShield recognises (emails, phone numbers, national IDs, IBANs and the names of people in your organisation) is pseudonymised, and the mapping lives in memory only. We do not claim that no personal data leaves: it depends on what it recognises. The risk analysis of supplier scans does not pseudonymise: it sends the supplier's domain and the findings.

Is it a RAG system?

No. It uses tool calls over your organisation's live data at the moment you ask, not an indexed document store.

Does the AI write the Virtual CISO reports?

No. They are deterministic. Only the advisory panel uses AI.

Can I switch it off, and how long does it keep things?

An administrator can switch it off at any time; while off, no call is made to the model. Conversations are kept for 90 days at most and deleted automatically.

See HodeiAI propose and a person decide

In the demo you ask it a question, see the proposal with its single-use approval and review the action ledger.

HodeiAI proposes; it never executes without a person.