beenuar/AiSOC

AiSOC receives telemetry data from various security devices, normalizes the format, generates alerts using a detection rule library, correlates alerts into incidents by entity, and then has AI agents investigate and record their reasoning process in an investigation ledger. Proposed remediation actions are only sent to vendor systems after human approval. New threat intelligence is retroactively applied to previously collected historical data. By default, it uses Ollama to run a quantized llama3.2 3B model of approximately 2 GB for triage, but you can also switch to a GPU, a model on the host machine, or your own service provider.

AiSOC receives telemetry data from various security devices, normalizes the format, generates alerts using a detection rule library, correlates alerts into incidents by entity, and then has AI agents investigate and record their reasoning process in an investigation ledger. Proposed remediation actions are only sent to vendor systems after human approval. New threat intelligence is retroactively applied to previously collected historical data. By default, it uses Ollama to run a quantized llama3.2 3B model of approximately 2 GB for triage, but you can also switch to a GPU, a model on the host machine, or your own service provider.

Why it matters The agent only calls typed tools and does not write query statements itself; every prompt and tool call is logged and auditable. It includes a local small model, so it can run without cloud API keys.

Stars 8229 (+31 / 24h)

Reasons

  • 31 new stars in 24 hours, total stars 8229 (indicating attention only); MIT License
  • The README explicitly lists "which features actually work" and clarifies that the included triage model is not a state-of-the-art model

Usage

  • After running git clone https://github.com/beenuar/AiSOC, execute make up in the directory (requires Docker Compose v2, 8 GB of RAM, and 20 GB of disk space); if you do not install Docker, you can first run pip install aisoc-sandbox and then execute aisoc-sandbox demo to run an offline walkthrough

Risks

  • The README admits that hosted model service providers have never been tested within the project, and the corresponding evaluation column states "Not tested"
  • Before connecting to real security devices and enabling automated remediation, you must verify the rules, permissions, and three safety switches yourself

Open

Trendshift(GitHub Trending 快照)

Explore the latest AI Radar updates