AiSOC: The 8,169-Star Open-Source AI Security Centre That Runs With Zero API Keys

AiSOC, an open-source AI security operations centre on GitHub, has passed 8,169 stars — up from about 6,353 in a mid-November roundup, a gain of roughly 1,800 stars in days — and it is climbing because of one unusual promise: it runs on a local language model with no API keys at all. Built by the developer known as beenuar, it ingests security telemetry, runs thousands of detection rules, correlates alerts into incidents, and investigates them with AI agents whose every step is recorded. The count was read live on github.com on 18 November 2026.

A security operations centre, usually shortened to SOC, is the team — or the software — that watches an organisation's computers around the clock for signs of attack, sorts real threats from noise, and decides what to do about them. AiSOC automates the middle of that job: the triage, correlation and investigation that normally eats a human analyst's whole shift. A detection rule is a written instruction that says "if you see this pattern in the logs, raise an alert".

What AiSOC does

The pipeline has four stages. First, telemetry arrives from the organisation's security tools and is normalised into a common shape. Second, the system runs its detection rules — 2,966 executable rules drawn from a library of 7,408, plus 76 sliding-window rules that watch for patterns across events. Third, alerts that belong together are grouped into incidents. Fourth, an AI agent investigates each incident and writes its reasoning to a replayable Investigation Ledger, then proposes an action that a human approves.

That last part is the design's most interesting choice. Every prompt and tool call the agent makes is recorded, so an investigation can be replayed step by step afterwards. New threat intelligence also re-sweeps the history already collected, looking for sightings of threats that were only recognised later. And a hypothesis can become a hunt without anyone writing a query: the model fills in a closed schema, so it cannot express a query of its own invention.

Analysts monitoring threat dashboards in the DNV Cyber security operations centre
Credit: DNV Cyber, via Computerworld Norway


AiSOC connects to real data in two ways. It can pull from 97 click-and-connect data connectors — Splunk, Sentinel, Elastic, CrowdStrike, Okta, AWS and Kubernetes audit among them — or it can accept pushed events through an ingest API. There are also four collectors for the surfaces a real cloud estate emits to: a CloudTrail trail on S3, a Google Cloud Pub/Sub subscription, an Azure Event Hubs capture, and a syslog listener.

The numbers behind the surge

FigureValue, as read 18 Nov 2026
GitHub stars8,169
Forks836
Executable detection rules2,966 of a 7,408-rule library
Data connectors97
Commits2,127
Releases51 (v23.5.1 in the README)
Created2 May 2026
LicenceMIT License

The growth is real but needs context. A star count measures interest, not deployment: 8,169 people thought the project worth bookmarking, which is not the same as 8,169 organisations running it in production. The MIT licence is permissive, so there is no licence catch for using or modifying the code. But the project is six months old, and its own README is unusually honest about that. It says "Executable is earned, not declared": a rule only joins the compiled ruleset after a vendor-shaped event has been replayed through the real connector and engine and the rule has been watched to fire. Two rules still cannot fire, down from 119.

No API keys — and what that costs

Running with no API keys is AiSOC's headline feature, and it works because the triage runs on a bundled local model. The first run downloads a model of about 2 GB, and the documented requirements are Docker Compose v2, 8 GB of memory and 20 GB of free disk. Setup is short — clone the repository, run the installer, and prove it works with a built-in smoke test that posts one real event through the pipeline and reports pass or fail at every stage.

It also ships an MCP server for Claude, Cursor and Continue. MCP, the Model Context Protocol, is the open standard that lets an AI assistant talk to outside tools and data sources through a common plug. AiSOC's MCP server means an assistant such as Claude can query the security centre directly, rather than the analyst copying alerts between windows.

The trade-off is honest and practical. A local model means no per-token bill, no data leaving your network, and no dependency on a vendor's uptime — but the machine doing the triage has to be powerful enough, and the model is smaller than the frontier models a cloud service would use. Google exploit_agent showed the other side of the same coin: AI agents are now being used to attack as well as defend.

Human approval is the default, not an option

The README's maturity section is the part a careful reader should spend time on. Playbooks triggered by an alert preview before they act: three switches must agree — the deployment, the tenant and the playbook — and every default is off. An approval step is a durable pause: the run suspends, survives a restart, resumes after the approval, and expires with a recorded outcome rather than hanging.

That caution is the right response to the risk that makes security teams nervous about AI agents: an agent that can act on its own can also act wrongly, at machine speed, across production systems. The Cursor Claude Code extension malware episode showed how easily a trusted tool can become an attack path, and AiSOC's answer is to make every automated action earn its execution.

A corridor between rows of server racks in the Virginia Tech data centre
Credit: Christopher Bowns, via Flickr


The project also refuses to pretend a dashboard is a deployment. A tenant with nothing connected lands on a setup wizard rather than an all-zero dashboard, and the console says so rather than showing placeholder numbers. Sample data can be loaded to watch five scenarios travel the same ingest path a real connector uses.

Why it is trending now

AiSOC is not a 17 November launch, and it should not be presented as one. It was created on 2 May 2026 and has been built in public since, with 51 releases and more than 2,100 commits. What changed is attention: the star count jumped by roughly 1,800 in a few days.

The audience it attracts will recognise the pattern from other breakout repos. Impeccable GitHub skill and the Hallmark GitHub skill both earned their stars by solving one concrete workflow well and documenting it honestly, and AiSOC follows the same playbook.

For a small security team, the honest pitch is this: AiSOC is a free, self-hosted starting point for AI-assisted triage, with a replayable record of every agent decision and a human approval gate on every action. For a large one, it is a reference design worth reading even if it is never deployed.

Analysts at workstations beneath a Cyber Fusion Center security-events wall display
Credit: Kudelski Security, via SKIDATA


Eight thousand stars say thousands of people want an open-source AI SOC that runs without API keys. Whether AiSOC becomes the tool they run every day will be decided by the next few months of releases — and by whether those two rules that still cannot fire finally do.

Key Takeaways

  • AiSOC (beenuar/AiSOC) is an open-source, self-hostable AI security operations centre with 8,169 GitHub stars as read on 18 Nov 2026, up about 1,800 in days.
  • It runs 2,966 executable detection rules from a 7,408-rule library, correlates alerts into incidents, and investigates with LLM agents whose every step is recorded in a replayable ledger.
  • It runs on a local model with no API keys, needs Docker Compose v2, 8 GB memory and 20 GB disk, and ships an MCP server for Claude, Cursor and Continue.
  • Actions need human approval by default; playbooks preview before acting, and the README says "Executable is earned, not declared" — 2 rules still cannot fire.
  • A star count is interest, not a deployment: AiSOC is trending, not newly launched — it was created on 2 May 2026.

Conclusion

AiSOC's surge says something simple about what security teams want right now: an AI SOC they can run themselves, on their own hardware, with no API key and a full record of what the agent did and why. The project's honesty about its own limits is the strongest reason to take it seriously. Stars are not deployments, and a trending repo is not a production system. But for anyone who has watched AI agents attack as well as defend, a free, auditable, human-gated SOC is worth an afternoon of testing.

Sources

  • beenuar/AiSOC, GitHub repository and README, read live 18 Nov 2026, 14:52 UTC: 8,169 stars, 836 forks, 321 watchers, MIT License, created 2 May 2026, 2,127 commits, 51 releases.
  • AiSOC README: 2,966 executable rules of a 7,408-rule library, 97 data connectors, Docker Compose v2 with 8 GB memory and 20 GB free disk, ~2 GB model download, MCP server for Claude, Cursor and Continue.
  • AiSOC README, "Project maturity": "Executable is earned, not declared"; 2 rules still cannot fire; playbooks default to preview and approval.

 

Comments