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.

3.9 KB
Raw Blame

Federated Threat Intel Mesh — threat model & design (v8 P1)

The mesh is opt-in gossip between self-hosted AiSOC instances via a lightweight, open-source hub (the community hub runs at mesh.tryaisoc.com; you can run your own). Its promise: every install makes every other install smarter — without any instance revealing its data.

This is the one feature class a closed SaaS vendor structurally cannot copy credibly, because trusting it requires reading the code. So the code is the product: services/mesh/.

What is shared

Exactly two artifact types, both aggregate and privacy-preserving:

  1. IOC sightings. SHA-256(normalized "<type>:<value>") + a coarse type (ip/domain/hash/url) + severity + first/last-seen. The raw IOC is never published. A peer learns the value only if it already has it (it can compute the same hash) — a private-set-intersection style exchange: querying returns a reputation only for a hash you supply, and only once k-anonymity is met.
  2. Verdict signatures. The institutional-memory signature key (category + connector + primary technique) plus a verdict distribution and mean confidence. No tenant data, no entities, no free text.

What is never shared

Raw IOC values, entity names, hostnames, usernames, IPs, tenant identifiers, alert descriptions, or any per-alert free text. mesh preview prints the exact outbound payload before you enable anything.

Privacy gates (defense-in-depth)

Gate Enforced How
k-anonymity hub-side + client A signature/sighting's consensus is revealed only when >= k distinct instances report it (default k=5, AISOC_MESH_K). Below k, a query is indistinguishable from "unknown".
Per-instance signing hub-side Every artifact is Ed25519-signed; the hub verifies before counting. One actor can't inflate consensus with sock-puppet IDs — each distinct instance is a distinct verified public key (see test_same_instance_cannot_inflate_consensus).
Opt-out client + hub Tenant-level and rule-level opt-out on the client; the hub also honors a per-instance opt-out (refuses to accept or serve).
Outbound audit client + hub Every shared artifact is logged; the hub keeps a per-instance receipts log.
Preview before enable client mesh preview shows exactly what would be shared.

Consumption — the mesh.py verdict stage

The verdict engine gains a deterministic mesh stage (services/mesh/app/consensus.py:mesh_contribution). Community consensus for an alert's signature produces a bounded verdict delta capped at ±0.10 — the mesh can nudge, never dominate, a local verdict. High community FP-rate pulls a verdict toward benign; overwhelming TP confirmation nudges it up; below k-anonymity it contributes nothing. The cap is unit-tested (test_mesh_contribution_is_bounded_and_directional). The UI shows a community consensus chip ("community: 41 instances saw this signature, 96% FP").

Measuring the lift (honesty note)

The claim "the mesh reduces false positives" must be measured, not asserted. The intended gate compares the eval harness with the mesh stage enabled vs. disabled and publishes the delta on the benchmark page. Until that longitudinal measurement is run on real multi-instance data, the benchmark page must present any mesh lift as synthetic/simulated and labelled as such — never as a real measured production number. (Same synthetic-vs-real rule as the rest of the eval harness.)

Trust posture

  • The hub is open source and runs on your infrastructure if you don't trust the community hub.
  • The hub only ever sees hashes and aggregates; it cannot reconstruct your environment.
  • Public feed intel is intentionally global; the mesh targets tenant-private signal, which is why sharing is opt-in and gated.

See SECURITY.md for the mesh disclosure policy.