AiSOC Open-Sources AI Security Operations Center With Decision Ledger

What Happened

AiSOC, an open-source AI security operations center, shipped with four integrated capabilities: alert fusion, LLM-agent triage, MITRE ATT&CK-aligned investigation, and a replayable decision ledger that records each agent step. The project is self-hostable, runs without external API keys, carries an MIT license, and ships an MCP server compatible with Claude, Cursor, and Continue. It gained 714 GitHub stars in a single day, per the source repository at github.com/beenuar/AiSOC.

Why It Matters

The decision ledger is the operationally consequential piece. Security agents that triage alerts and take investigation steps are difficult to deploy in regulated or audited environments without a verifiable record of what each agent did, in what order, and on what evidence — a replayable log converts agent behavior from an opaque runtime artifact into an inspectable audit trail. This directly addresses the adoption blocker for agentic security tooling inside enterprises with compliance obligations, where "the model decided" is not an acceptable answer to an auditor. The absence of required API keys lowers the barrier to evaluation: teams can stand up the system and inspect its behavior before committing to model vendors or inference budgets. MIT licensing removes procurement friction for internal forks. The combination positions AiSOC as a reference implementation for auditable agent security operations rather than a product.

Technical Details

The system fuses alerts from multiple sources into a unified queue, then routes them through LLM-agent triage with investigation steps mapped to MITRE ATT&CK techniques — giving each finding a structured vocabulary rather than free-text reasoning. The decision ledger records every agent step in a replayable format, meaning an operator can reconstruct the exact sequence of triage decisions and evidence seen at each point. The bundled MCP server exposes the system to Claude, Cursor, and Continue, so agent clients can query and act on the SOC through the Model Context Protocol rather than bespoke integrations. Keyless operation implies the system ships with local or bundled inference paths, though the practical ceiling on triage quality under that configuration versus frontier models is the open question. Self-hosting means operators own the data plane, but also own scaling, storage growth from ledger retention, and model lifecycle management. The 714-star single-day velocity reflects discovery, not production validation — no benchmark or deployment figures were included in the signal.

Operational Impact

For security teams, the immediate change is evaluability: a self-hosted, keyless deployment can be running against real alert data in hours, with the ledger providing the evidence needed to justify or reject agent recommendations before granting any autonomy. Detection engineers gain a replay primitive — when an agent mis-triages, the ledger supports root-cause analysis of the decision path rather than guesswork, which shortens the iteration loop on triage logic. The MCP server collapses integration work for teams already using Claude, Cursor, or Continue as their agent surface; SOC data becomes queryable through the same protocol as other tooling instead of requiring custom connectors. Ledger retention introduces a new storage and privacy obligation: decision records containing alert context inherit the sensitivity of the underlying telemetry and must be governed accordingly. For vendors selling closed AI-SOC platforms, the competitive pressure shifts from feature parity to auditability parity — a replayable ledger becomes table stakes rather than differentiation.