AiSOC — Open-Source Self-Hosted AI Security Operations Center
Open-source, self-hostable AI SOC with alert fusion, purple-team drills, agent-assisted triage, and MITRE ATT&CK investigation.
Evidence shows: README explicitly states default copilot/read-only mode, autonomous response requires high confidence and reversible low-blast radius, with rollback mechanism; data flow docs detail pseudonymization and local model options; dependencies have pnpm overrides for known vulnerabilities; external effects have clear policies (e.g., mesh is opt-in and privacy-preserving). Deductions: publisher identity unverified, and no execution evidence, so some criteria only partially met.
Evidence shows: README claims CI gating and tests, but no test results provided; dependency availability has pnpm lock and overrides, but not verified; failure messages documented but not actually tested. Deductions: static review cannot confirm actual reliability.
Evidence shows: README provides multiple installation paths (Docker, Codespaces, Python, etc.), rich scenarios (multiple bundled scenarios), good environment fit (supports multiple deployment platforms). Deductions: capability boundaries described clearly but not fully verified.
Evidence shows: clear information architecture (docs directory, architecture diagrams), detailed install notes, stable naming (version numbers), rich examples and FAQ, known limitations mentioned, MIT license explicit, CHANGELOG and versioning present, maintenance responsibility via SECURITY.md and contributing guide. Deductions: maintenance responsibility not explicitly assigned to a person or team.
Evidence shows: output usability has JSON output and demo, high marginal value (open-source AI SOC), good cost-benefit (free self-hosted). Deductions: not actually run, so output quality unverified.
Evidence shows: README provides many doc links and CI gating claims, but no actual test results; cross-source verification limited; fact-inference separation good (REALITY_REPORT). Deductions: static review cannot verify actual tests.
- Publisher identity unverified; proceed with caution.
- Static review cannot confirm actual runtime behavior; test in an isolated environment.
- Dependency security has overrides but not all dependencies verified.
What does this agent do, and when should you use it?
AiSOC is an open-source, self-hosted AI Security Operations Center (SOC) released under the MIT license. It integrates security event ingestion, correlation, AI-driven investigation, and a unified SOC console. Its core is a ~600-line LangGraph orchestrator that is readable and patchable. Every step of the agent (prompts, tool calls, rationale) is logged in the Investigation Ledger, which is replayable. It ships with 83 click-and-connect data connectors, 947 executable detection rules (869 native), and a 6,000-rule provenance-tracked imported library. Deployment is documented for Docker Compose, Kubernetes, and Terraform, and it provides three quickstart paths: a Python-based aisoc-sandbox simulator, in-browser GitHub Codespaces, and a Docker-based install script for the full environment. It includes an MCP server for IDE integrations and a public CI-gated benchmark harness for evaluation.
AiSOC operates as an end-to-end security operations platform: it ingests events via connectors (Python, 78 vendors) or osquery-tls, normalizes them to OCSF in a Go ingest service, enriches with IOC and Shodan, and publishes to an Apache Kafka event spine. Detection engines (Sigma, YARA, KQL) process the stream through Fusion (ML), UEBA (baseline), and a rule engine, with fused alerts. The LangGraph agent performs autonomous investigation in an LLM tool-calling loop with real tools (IOC enrichment, MITRE lookup, graph blast radius). Results are recorded in the Investigation Ledger and surfaced in a web console with an /explore data explorer, a /hunt natural-language workbench, and a marketplace for plugins and detections. Response execution is policy-governed with copilot mode as default and prompt-injection guarding. The platform also offers federated search (Splunk SPL, Sentinel KQL, Elastic ES|QL, QRadar AQL), stateful/windowed detections (brute-force, password-spray, port-scan), and cost-governed LLM routing with per-tenant budgets and circuit breakers.
- Security teams wanting an auditable, replayable AI investigation process with self-hosting for compliance.
- Security engineers who need to validate alert reduction with public eval tooling and CI-gated benchmarks.
- Analysts who want to write natural-language hypotheses and get ES|QL / SPL / KQL queries back from the /hunt workbench.
- Investigators who want to explore the event lake from a single surface with pivots to identity/graph/intel via /explore.
- MCP users who want to manage alerts and replay investigations from within IDEs like Claude, Cursor, or Cody.
- Purple teams leveraging AiSOC's detection-as-code lifecycle and automated response policies for security drills.
What are this agent's strengths and limitations?
- MIT license and self-hostable with no vendor- cloud callbacks; fully air-gapped possible with local models (Ollama/vLLM) and pseudonymized evidence for hosted LLMs.
- Complete audit trail: the Investigation Ledger records every prompt, tool call, and evidence chunk, replayable and shareable as redacted public permalinks.
- Strong evaluation culture: public CI-gated benchmarks for alert reduction and substrate self-consistency checks, backing product claims (46 gated).
- Highly extensible: plugin SDK (Python, TypeScript, Go), detection-as-code, MCP server, and a three-model (Semantic, Behavioral, Knowledge) architecture.
- Cost-governed LLM routing with per-tenant budgets, content-addressed caching, and multi-model gateway fallbacks.
- Installation complexity: the full local stack requires Docker, pnpm, and Node.js, and manages many services (Kafka, ClickHouse, etc.).
- Some components are still marked preview or planned for v8.0: e.g., the
npx aisocCLI is not yet available, and the MCP server is not yet published to npm. - Autonomous response is limited to copilot/read-only by default, so full automation requires policy configuration.
- Detection count (947) may be smaller than some commercial SIEMs, though a 6,000-rule imported library exists.
- Some screenshots and videos in the docs are placeholders, which may hinder UI evaluation.
How do you install or deploy this agent?
Several installation paths are documented, depending on what you have:
- Zero-dependency quickstart (Python 3.10+, no Docker):
pip install -e packages/aisoc-sandbox
aisoc-sandbox demo
- Full local stack (Docker and pnpm):
git clone https://github.com/beenuar/AiSOC && cd AiSOC && pnpm aisoc:demo
- Bootstrap on a clean machine:
curl -fsSL https://raw.githubusercontent.com/beenuar/AiSOC/main/install.sh | bash
- Production deployment: refer to
apps/docs/docs/installation.md(Render, Fly.io, Docker Compose, Kubernetes, Terraform). Helm charts are ininfra/helm/, Terraform ininfra/terraform/.
Note: npx aisoc is not available until v8.0; today it builds from packages/aisoc-lite/.
How do you use this agent?
After the quickstart (full stack), open http://localhost:3000 in your browser to see the console. The console shows the seeded LockBit 3.0 case (INC-RT-001) with the agent's investigation streaming prompts, tool calls, and rationale in the Investigation Ledger. Stop the stack with pnpm aisoc:demo:down.
For production usage: configure connectors (via console or code), ingest events, and detections/alerts will appear. Use /explore for data exploration and /hunt to create hunt queries in natural language. The MCP server (services/mcp/) exposes 13 tools for direct IDE usage.
CLI quickstart (should be available after v8.0):
npx aisoc triage --demoHow does this agent compare with similar options?
The README compares AiSOC against Wazuh, Splunk ES, and closed-source AI SOCs. Key advantages include a more permissive open-source license (MIT vs GPL-2 or proprietary), autonomous AI investigation, a public Investigation Ledger, CI-gated evaluations, and full self-hosting capabilities.
FAQ
Does AiSOC call external APIs?
What credentials are required to run AiSOC?
How does AiSOC protect against prompt injection?
Is multi-tenancy supported?
Is it difficult to add custom detections or connectors?
detections/ with positive/negative fixtures. Connectors are implemented by subclassing BaseConnector and registering. CI validates these changes on every PR.