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.
Data handling by deployment mode
What leaves your perimeter, what is stored, and for how long — answered per mode, because the answers genuinely differ and a single answer for all of them would be wrong three times out of four.
docs/trust/data-flows.md is the short version of
the egress question. This page is the whole data lifecycle.
The four modes
| Mode | How it is started | Who runs the infrastructure | Who runs the model |
|---|---|---|---|
| Self-hosted, bundled model | make up |
You | You — ollama in your own stack |
| Self-hosted, hosted model | make up, then configure a provider in the console |
You | A third party you choose and contract with |
| Air-gapped | infra/compose/docker-compose.airgap.yml |
You | You, with no route off the host |
| Managed | Operated for you | The maintainers | See sub-processors |
The default is the first. make up needs no account, no key and no GPU: it
ships llama3.2:3b-instruct-q4_K_M and triage produces a real verdict from a
real model with real token counts. That matters for this page because the
default configuration makes no model call off your infrastructure at all.
Egress, per mode
What never leaves, in every mode
- No telemetry to the AiSOC project. There is no phone-home and no
"model improvement" callback. This is enforced, not asserted:
scripts/check_default_egress.pyreads every service's declared settings defaults and fails on a public host that no air-gap guard covers — 25 URL defaults across 11 settings modules, 23 private and 2 allow-listed with a reason.tests/test_no_default_egress.pyputs a socket guard in front of each service's import and ASGI startup, because the static gate cannot see a URL built at run time. - Your stores. Postgres, Kafka, Redis, and the optional ClickHouse, Neo4j and Qdrant are containers in your own stack. Nothing replicates them outward.
A third control covers what the first two structurally cannot: the static gate
is blind to run-time URLs and the socket guard is blind to paths startup does
not reach. container-egress.yml
starts images built from the commit under test on a Docker network created
--internal, with a DNS sinkhole as their only resolver, and classifies every
name they ask for. It carries a canary — a container wired identically that
must be seen by the sinkhole — because zero observations is what a correct
run and a blind probe both look like.
Self-hosted with the bundled model
Nothing leaves for triage. The model runs in the ollama container beside
everything else.
Two things still leave if you enable them, both on purpose and both yours to turn off:
- Connector polling reaches the vendors you connected. That is the product.
- Threat-intel feeds you enable. The one that is on by default is the CISA Known Exploited Vulnerabilities catalog — a public, authoritative feed that needs no API key and receives nothing from you but the request.
Self-hosted with a hosted model
Evidence reaches the provider you configured, pseudonymized.
This is the paragraph that was wrong for a long time, so it is worth being
exact about what changed. The reversible pseudonymizer
(services/agents/app/privacy/redactor.py)
existed and was unit-tested, and no LLM call site invoked it — so the
documentation described a planned control as a shipped one. It is now applied
at the contract layer
(services/agents/app/llm/contract.py
calls egress_privacy.open_session), which is the one place all sixteen agent
call sites already pass through.
| What is replaced | With |
|---|---|
| Internal IPs, internal hostnames | IP_3, HOST_2 |
| Usernames, email addresses | USER_1 |
| File paths, secrets | opaque per-run tokens |
| Public threat indicators | nothing — preserved deliberately, or the agent cannot reason about them |
Tokens are per-run and in-memory. The model reasons over tokens; the ledger and console re-hydrate real values locally.
The gate is test_egress_pseudonymization.py,
and it is worth noting how it is written: it drives safe_ainvoke and
inspects what a fake provider actually received, rather than calling the
redactor directly. Testing the path rather than the function immediately found
a real gap the unit test could not — a bare username in prose
(running as priya.raghavan) went out in the clear, because usernames were
only redacted in the DOMAIN\user form or under a user-ish key.
You can turn it off. If you do, raw evidence reaches the provider. Use that only under a signed zero-retention agreement.
Air-gapped
infra/compose/docker-compose.airgap.yml ships the local model and expects no
route off the host. This is the only mode in which "no evidence leaves" is
unconditionally true, and it is the mode the container-egress job exercises.
Managed
The maintainers operate the infrastructure. See sub-processors, which says plainly what this repository can and cannot establish about that arrangement.
What is stored, and where
| Data | Store | Notes |
|---|---|---|
| Alerts, cases, users, tenants | Postgres | Row-level security plus query-layer predicates |
| Connector credentials | Postgres, as ciphertext | Never in the clear; see the vault |
| Audit log | Postgres | Hash-chained, and a BEFORE UPDATE OR DELETE trigger (migration 004) refuses both |
| Investigation Ledger | Postgres | Every agent step, with the prompt and response |
| Raw events | ClickHouse | Full profile only. Absent on a default install |
| Entity graph | Neo4j | Full profile only |
| Embeddings | Qdrant | Full profile only |
| Cache, queues, rate limits | Redis | Tenant-prefixed keys |
| In-flight events | Kafka | Tenant in the envelope; retention is your broker's setting |
Encryption
| State | Where | |
|---|---|---|
| Connector credentials | Fernet AES-128-CBC + HMAC-SHA256 by default (vault:v1:), or envelope encryption with a KMS-held key (vault:v2:) |
credential_vault.py, envelope_cipher.py |
| Backups | AES-256-GCM, chunked with a per-chunk nonce and additional authenticated data binding the chunk index and the final flag, so a reordered or truncated archive is refused | scripts/backup_crypt.py |
| Everything else at rest | Your disk encryption | A property of your infrastructure, not of this software |
| In transit, browser to edge | Your TLS terminator | Not shipped; the compose stack serves plain HTTP and expects a reverse proxy |
| In transit, service to service | Plaintext inside the container network on the default compose stack | Stated rather than implied. Deployments needing mutual TLS between services should supply it at the platform layer |
The backup row is specific because it used to be false: the script gzipped and
uploaded while the documentation claimed AES-256-GCM. Two things make the claim
real now. openssl enc refuses AEAD ciphers, which is why the encryption
is a Python script on cryptography rather than a shell one-liner. And
Backup → destroy → restore (Postgres via S3) is a required status check
that seeds real data, encrypts, destroys, restores, measures recovery time and
proves a bit-flipped archive is refused.
Retention and deletion
- Per-tenant retention is implemented in
services/api/app/services/retention.py. The purge worker ships default off and dry-run first, deliberately: a deletion job that starts deleting on upgrade is a worse failure than one that has to be switched on. - Tenant deletion is
services/api/app/services/tenant_deletion.py. It exists because 14 tenant-scoped tables had no foreign-key cascade, so deleting a tenant orphaned institutional memory and compliance evidence rather than removing it. - Kafka retention is your broker's configuration, not ours.
- The audit log is append-only by design, which is in tension with a deletion request and is the right tension to have. The hash chain is what makes the log evidence; a log that can be edited to satisfy a request is not one.
Subject-access and portability
There is an audit export surface
(services/api/app/services/audit_export.py)
producing CSV and HTML bundles, and a report builder for case material.
What does not exist is a dedicated data-subject-request endpoint. ADR-0002 records the intent and its open question; no such route is in the tree today, and the questionnaire lists it as a gap rather than describing it as a feature.