Skip to content

armgit / AiSOC

Public Imported from GitHub
0

Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.

5.9 KB
Raw Blame

Sub-processors

Status: incomplete, deliberately. There is no verified sub-processor list for the hosted service in this repository, and this page does not invent one. What follows is (1) the answer for self-hosted deployments, which is complete and checkable, and (2) a clearly-marked placeholder for the hosted service, with exactly what is missing and who can supply it.

Self-hosted: there are none, by construction

If you run AiSOC yourself, the project processes none of your data and operates no infrastructure in the path. There is no phone-home, no telemetry to the AiSOC project, and no model-improvement callback. That is enforced by three independent controls rather than asserted — scripts/check_default_egress.py, tests/test_no_default_egress.py and container-egress.yml — and the detail is in data handling.

The processors in a self-hosted deployment are therefore the ones you choose:

You choose It receives Your control
Your cloud or hardware Everything Yours entirely
Each vendor you connect API calls from the connector you configured, under the credential you issued Least-privilege scopes
A model provider, only if you configure one Pseudonymized evidence, or raw if you turn pseudonymization off Per-mode egress
Each threat-intel feed you enable A request. The only default is the public CISA KEV catalog, which receives nothing from you Disable it

A contract with any of those is between you and them. The project is not a party to it and cannot be.

Hosted service: what this repository cannot establish

The maintainers operate a managed offering. A sub-processor list for it is a legal artefact describing the processing arrangements actually in force, and it cannot be derived from source code. This page will not guess at one.

What the repository does say, and why that is not an answer

Two documents describe hosted infrastructure:

Three reasons that is documentation and not a sub-processor list:

  1. A deployment description is not a processing agreement. A sub-processor list states who processes personal data, for what purpose, in which jurisdiction, and under which contract. None of that is in a compose file or a Terraform module.
  2. It cannot be verified as current from here. Infrastructure changes without a commit, and a hosting arrangement that moved would leave these pages reading exactly as they do now. A reader has no way to tell a current page from a stale one, and neither does this one.
  3. The list is certainly longer than the hosting vendor. Payment processing, transactional email, error tracking, analytics and support tooling are all plausible, each is a real sub-processor if present, and naming any of them without confirmation would be the fabrication this page exists to avoid.

The gap, stated

ADR-0002 committed to publishing a GDPR posture at docs/compliance/gdpr.md — controller/processor mapping, a DPA template, a data-subject-request workflow and a sub-processor list — "alongside this ADR".

Neither docs/compliance/gdpr.md nor docs/compliance/README.md exists in the tree. The single occurrence of the word "sub-processor" anywhere in this repository is the ADR line promising the list. That is recorded here as an open commitment rather than quietly dropped, which is the whole point of checking an ADR's deliverables against the tree.

What has to happen, and by whom

This is a maintainer action, not a code change. Completing it needs:

  • The current hosting, database and cache providers for the managed service, confirmed rather than inferred from infra/.
  • Every other vendor that touches customer data in that path — payments, email, error tracking, analytics, support.
  • For each: purpose, data categories, processing location, and the DPA or standard contractual clauses in force.
  • A change-notification commitment, since a sub-processor list without one is a snapshot rather than a control.
  • The GDPR posture document ADR-0002 committed to, of which this list is one section.

Until those are supplied, a prospective customer of the hosted service should ask for the list directly and not infer it from this repository. Self-hosting removes the question entirely, which is the honest recommendation while this page is in the state it is in.

Models are the sub-processor people forget

Worth separating out, because it is the one most often missed in an AI product's diligence.

On the default self-hosted install there is no model sub-processor at all. make up ships llama3.2:3b-instruct-q4_K_M and runs it in your own stack — no account, no key, no GPU. Triage produces a real verdict from a real model with real token counts.

A hosted model provider becomes a sub-processor only when you configure one, and then it is your contract. What it receives in that case is pseudonymized at the contract layer unless you turn that off; the mechanism, the token scheme and the deliberate exception for public threat indicators are in data handling.