Neural Ghost · Platform Architecture
Ghost SwarmA next-generation autonomous security platform: it fuses external
intelligence with internal exposure into one attack graph, computes the routes to your
crown jewels, and validates exploitability in an authorized lab — without touching production.
from “here are your vulnerabilities” to “here is the path an attacker walks to your crown jewels —
and it starts with a credential already leaked on the dark web.”
0
mypy · ruff · bandit-high
01
The complete map
external intel → core → internal exposure
02
The action lifecycle
the choke point — no bypass, no shell strings
AGENT
Proposal
named tool + level + target — agents never execute directly
POLICY
Decision
deny-by-default; only an approved decision proceeds
EXECUTOR
Run
typed handler, or fail-closed for exploitation
CONTAINER
Isolate
active tools run with no egress, no caps, read-only, in-lab
EVIDENCE
Commit
SHA-256 tamper-evident chain — reproducible audit
03
Capability map
vs. the 57-section brief · all shipped
✓ Multi-agent orchestration
11 specialists + reactive, guarded event bus
§5–7 ✓ Asset knowledge graph
NetworkX, typed weighted edges
§8 ✓ Attack-path engine
enumeration + explainable 6-factor risk
§9 ✓ OSINT / dark-web correlation
leaked cred ⋈ exposed service ⋈ CVE → path
§10·53 ✓ Reasoning layer
heuristic / Ollama / DarkForensic-7B — proposes only
§11–12 ✓ Policy engine + evidence
deny-by-default · SHA-256 chain
§13–14 ✓ Code & supply-chain (GhostCode)
7 scanners + AI-code analyzer + dedup + SBOM
§15–16 ✓ Live attack-surface management
Docker · Kubernetes · Cloud collectors
§17–20 ✓ Interactive dashboard
self-contained SVG attack graph, nonce-CSP
§21 ✓ API + job system
FastAPI · in-process or RQ/Redis backend
§22–23 ✓ Observability
JSON logs · Prometheus /metrics · Grafana
§24 ✓ Self-security & sandbox
RBAC/OIDC · SSRF/XSS-hardened · isolated runner
§25–26·31–33
04
Security model
defense in depth · authorized-by-construction
Access API keys with roles (viewer · operator · admin) or
OIDC/JWT from an external IdP. Non-admin principals confined to their tenant.
Authorized offense Exploitation is off by default , gated to
lab targets , and confined to a container with no egress and no capabilities.
Non-destructive The exploit-runner confirms a weakness and
redacts any sensitive value — no exfiltration, no pivot, no persistence.
Hardened API Security headers, nonce-CSP dashboard, rate limiting,
SSRF-guarded ingestion, timing-safe auth, path-confined scans.
Self-audited bandit + pip-audit + manual review found and fixed a real
SSRF and stored XSS in our own code.
Prompt-injection safe External content is data, not instructions.
The model proposes allowlisted actions; the Policy Engine still approves them.
05
Runtime & deployment
local-first · scales up when asked · self-hosted
Storage SQLite by default (offline / air-gap);
PostgreSQL verified live for multi-node.
Jobs In-process by default; RQ + Redis workers for
horizontal scale — verified running.
Reasoning Offline heuristic default; Ollama or the production
DarkForensic-7B model when configured.
Deploy docker compose up — hardened container;
profiles for postgres · queue · monitoring · labs .
Continuous ASM ghostswarm-asm CLI + a systemd
timer keep the attack surface current.
Quality gates ruff + mypy clean (both CI-enforced);
200 tests; bandit 0-High; pip-audit clean.
Everything above verified running in Docker on a real host —
containers, a live k3s cluster, Postgres, a Redis worker, Grafana.
self-hosted · docs/ ARCHITECTURE · adr/ · SECURITY · THREAT_MODEL · STATUS