AiSOC: a free, self-hosted AI security operations center that logs every step its agents take

AiSOC, an open-source "AI Security Operations Center," was one of GitHub's fastest-rising security projects on November 2: a daily trending tracker logged 446 new stars that day, taking the Python project to roughly 6,400. It promises something security teams usually pay a lot for — AI agents that triage and investigate alerts — in a package you host yourself, under the MIT license.

Why it matters: "AI SOC" products are one of the hottest pitches in commercial security, and most are closed, hosted services that hold your logs. AiSOC's README says it keeps data in your own deployment, records every prompt and tool call its agents make, and never responds to an incident without a human approving it. That makes it a useful way to see — and audit — what an AI analyst actually does.

What it does

According to the README, AiSOC runs a pipeline from ingest to investigation:

• Detection and alert fusion. Normalised telemetry runs through 2,530 executable detection rules (out of a 7,015-rule library on disk) plus 76 sliding-window rules, and related alerts are grouped into incidents.
• AI triage and investigation. Agents read alerts and context, call typed tools and write prompts, tool calls, citations, verdicts and token costs to an Investigation Ledger, exportable as signed evidence bundles in the OCSF 1.9.0 format. The project describes MITRE ATT&CK investigation as a core feature.
• Retro-hunts. New threat intelligence re-sweeps stored history, using a library of 68 hunts; the CISA Known Exploited Vulnerabilities catalog is polled without an API key.
• Human-approved playbooks. Playbooks preview by default and pause for approval; the README says response "is not autonomous by default."
• A copilot that cites its work, labelling any claim it can't back with a ledger entry as "uncited."
• Connectors and migration. The README lists 87 connectors, including Splunk, Sentinel, Elastic, CrowdStrike, Okta, AWS and Kubernetes audit logs, plus a tool that translates Splunk SPL, Sentinel KQL and Elastic EQL rules.
• An MCP server. The project's FAQ says the @aisoc/mcp package exposes 11 Investigation Ledger tools to clients such as Claude Desktop, Cursor and Continue, so analysts can query cases and replay agent decisions from their editor.

Out of the box it uses a pinned ~2 GB Llama 3.2 3B model via Ollama on CPU. The FAQ says models are mapped in a single LiteLLM config, so you can point it at OpenAI, Anthropic, Azure OpenAI or a self-hosted open-weights server instead.

How to try it

From the README (needs Docker Compose v2 with 8 GB of memory and 20 GB of free disk, plus Python 3.9+ and bash):

git clone https://github.com/beenuar/AiSOC && cd AiSOC
make up            # core profile: 16 services, ~8 GB RAM
make demo          # optional: load labelled synthetic incidents

make up-full adds an event lake, entity graph, search and enrichment (22 services, ~12 GB). The README also offers a lighter sandbox: pip install aisoc-sandbox && aisoc-sandbox demo. It lists one-line curl | bash and PowerShell installers too — read any script before piping it to a shell, especially on a security box.

Limitations

The README is unusually candid about what isn't proven:

• No hosted LLM provider has been exercised by the project, and its benchmark figures are described as self-consistency measures, "not live agent accuracy."
• "Executable" doesn't mean "detects attacks." A rule counts as executable if replayed events can reach it; two still can't fire.
• Not a SIEM replacement. It doesn't do long-term retention or compliance search and can't see telemetry you haven't connected.
• Prompt injection isn't solved. The FAQ says its adversarial tests measure resilience but don't guarantee protection — a real concern for any agent that reads attacker-controlled log data.
• Privacy chores are manual. Retention is off by default (data is kept until you set a policy), and there is no dedicated data-subject deletion endpoint for GDPR or CCPA requests.
• Corporate backing. The README says development is funded by Cyble, and the FAQ floats a possible managed cloud tier later while the self-hosted version stays MIT.
• Docs and README figures don't always match (the FAQ cites slightly different rule counts), so check numbers against the version you run.

Who should care

Small security teams and MSSPs who want AI-assisted triage without shipping logs to a vendor, and anyone evaluating commercial AI SOC tools who wants a transparent baseline to compare against. Start with the demo profile and synthetic data, keep the human-approval gate on, and treat its verdicts as a second opinion rather than a replacement for an analyst.