What happened
AiSOC, an open-source Security Operations Center (SOC) built around AI agents, is now available on GitHub under the MIT license. Developed by Cyble, the tool collects security logs, correlates them into incidents, and uses AI agents to propose responses that require human approval. The project includes a local quantized Llama 3.2 model for CPU-only environments and features an 'Investigation Ledger' that records every prompt and tool call for auditability.
AiSOC is an open-source Security Operations Center that runs on user-owned servers, utilizing AI agents to triage security alerts. The project, initiated by Beenu Arora of Cyble, is licensed under MIT and requires Docker Compose v2 with at least 8 GB of memory and 20 GB of disk space. It normalizes telemetry from various security tools, applies over 2,500 detection rules, and groups alerts into incidents for AI-driven investigation.
The system uses a quantized Llama 3.2 model with 3 billion parameters served by Ollama, designed for CPU-only machines. While this model is not frontier-class, it allows for local, private inference. The project includes an 'Investigation Ledger' that logs every prompt, tool call, and citation, creating a signed, byte-for-byte identical bundle that allows auditors to retrace the AI's decision-making process without trusting a summary.
Safeguards are built around the model to prevent autonomous action. Verdicts citing missing evidence are downgraded to human review, and prompts are validated to reject raw logs or potential secrets. External actions require explicit human approval unless specific autonomy is granted. The project also includes an MCP server for integration with tools like Claude and Cursor, and a migration package that translates rules from Splunk, KQL, and EQL formats.
The project has faced security challenges, including a prompt injection vulnerability reported in May 2026 that allowed hostile text to trigger automatic alert closure. This was fixed within 15 hours through prompt sanitization. Additionally, the project published an internal audit in September 2026 acknowledging that its initial quick-start guide resulted in a non-functional setup with fabricated incidents, a flaw that has since been corrected in later versions.
Source details: pasqualepillitteri.it
Why it matters
AiSOC addresses the critical bottleneck of alert fatigue in security operations by automating the initial triage process with transparent, locally runnable AI agents. Its open-source nature and detailed audit logs provide a verifiable framework for integrating AI into defensive security workflows, allowing organizations to inspect exactly how AI decisions are made. This is significant for security teams seeking to adopt AI without sacrificing accountability or relying on opaque cloud services.
The primary value of AiSOC lies in its transparency and local-first architecture. By providing a readable, open-source codebase and a detailed audit trail of AI actions, it offers a practical test bench for security teams to evaluate AI agents without the black-box risks associated with proprietary cloud solutions. This is particularly relevant in an era where AI agents are increasingly deployed in sensitive security environments.
The tool addresses the 'triage bottleneck' in SOCs, where analysts are overwhelmed by repetitive alert sorting. By automating this step with a local model, AiSOC reduces the cognitive load on human analysts, allowing them to focus on complex investigations. The requirement for human approval on all external actions ensures that the AI remains a decision-support tool rather than an autonomous actor, aligning with current best practices for AI safety in critical infrastructure.
The project's frank acknowledgment of its own flaws, such as the initial broken quick-start guide and the prompt injection vulnerability, enhances its credibility. This level of transparency is rare in AI security tools and provides a realistic view of the challenges involved in deploying AI agents in production environments. It serves as a case study for the importance of rigorous testing and prompt sanitization in AI-driven security systems.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
Monitor the project's handling of prompt injection attacks, as demonstrated by a recent vulnerability fix, and observe whether the local 3-billion parameter model proves sufficient for complex threat detection or if users consistently migrate to larger, paid models. Also watch for the maturity of its automated playbooks, which currently require multiple manual switches to enable.
The effectiveness of the local 3-billion parameter Llama 3.2 model in real-world scenarios will be a key indicator of the project's utility. If the model proves too limited for complex threat detection, users may be forced to rely on larger, potentially cloud-based models, which could undermine the local-first privacy benefits. The project's performance metrics, such as the 44 usable answers out of 50 alerts before JSON constraint, suggest that further optimization may be needed.
The evolution of its prompt injection defenses will be critical. As AI agents become more integrated into security workflows, the risk of attackers using hostile text in logs or file names to manipulate the AI will increase. The rapid fix for the May 2026 vulnerability is a positive sign, but ongoing vigilance will be required to ensure that the system can handle the full spectrum of adversarial inputs.
The adoption of its automated playbooks, which are currently in preview and disabled by default, will determine whether AiSOC can move beyond triage to active response. If users find the multi-switch approval process too cumbersome, the tool may remain a passive analysis tool. Conversely, if the playbooks prove reliable, they could significantly reduce the time to respond to confirmed threats.