AiSOC
Open source (maintained by Beenu Arora, funded by Cyble) · AiSOC: Open-Source, Self-Hosted AI SOC With Auditable Triage and Investigation Agents
AiSOC is an MIT-licensed, self-hostable AI security operations center: it ingests telemetry from your security tools, runs detection rules, groups alerts into incidents, has AI agents triage and investigate each one with every prompt and tool call recorded in a replayable Investigation Ledger, and proposes responses that a human approves. It suits security teams, MSSPs and security engineers who want an AI SOC they can run, audit and modify on their own infrastructure, including air-gapped networks.
Best for
Security teams and MSSPs that want an AI SOC they can self-host, audit line by line and modify, including in air-gapped or regulated environments, and that have the engineering capacity to run a multi-service Docker or Kubernetes stack and choose their own models
Not ideal for
Teams that want a managed service with support contracts and SLAs, small IT teams without someone to operate Kafka, Postgres and a dozen-plus services, buyers who need proven verdict accuracy with a strong hosted model out of the box, and organizations that want a long-retention SIEM replacement
Who it's for
SOC analysts and managers, security engineers, detection engineers and MSSP operators who want to run an AI SOC on their own infrastructure
Capabilities
- End-to-end pipeline on one host: authenticated ingest, normalization onto a Kafka spine, rule-based detection, correlation into incidents, AI triage and a web console, brought up with `make up` and checked stage by stage with `make smoke`
- Four named agents (Detect, Triage, Hunt, Respond); Triage auto-classifies each alert and routes it to phishing, identity, cloud or insider analysis, and investigations run through a LangGraph orchestrator with recon, forensic, responder and reporter steps
- Investigation Ledger that records every prompt, model response, tool call with its inputs and outputs, citation, verdict and token cost, replayable step by step and exportable as a signed evidence bundle
- Grounding checks: a verdict citing an indicator that never appeared in the evidence is sent to human review instead of auto-closing, and prompts carrying raw logs or secret-shaped values are refused before they are sent
- Governed response: playbooks preview by default and execute only when the deployment, tenant and playbook all opt in, approval steps pause durably and survive restarts, and an approver must hold the right permission tier and cannot be the requester
- Per-action autonomy policy (auto, review, escalate or reject) for actions such as endpoint isolation, identity disablement, network blocks and email purges, with blast-radius limits and a global advisory-only stop
- 91 data connectors plus cloud collectors for CloudTrail on S3 via SQS, GCP Pub/Sub, Azure Event Hubs and a syslog/CEF listener, each with a resumable cursor
- Large detection library of native and imported rules (SigmaHQ, Splunk Security Content and others), where a rule counts as executable only after a vendor-shaped event replayed through the real connector and engine makes it fire
- Migration tool that translates Splunk SPL, Sentinel KQL and Elastic EQL rules and refuses rules it cannot carry rather than approximating them
- Hunting with a YAML hunt library, natural-language questions rendered to ES|QL, KQL and SPL, and retro-hunts that re-sweep collected history when new threat intelligence arrives (full profile)
- Bundled Ollama model (llama3.2 3B, about 2 GB) for CPU-only triage with no API key, and a LiteLLM gateway that maps each task alias (triage, investigation, copilot, hunt and others) to any hosted or OpenAI-compatible self-hosted model
- MCP server for Claude Desktop, Cursor, Continue.dev and Cody that lists alerts and cases, queries detections and the event lake, starts investigations, replays the agent's reasoning and previews (never performs) response actions
- Enterprise controls: SAML and OIDC sign-in, SCIM provisioning, console MFA, RBAC, multi-tenant row-level isolation, an append-only audit log, and a CISA KEV threat feed that works without a key
Limitations
- You run and maintain the stack yourself: the core profile is 16 services and needs Docker Compose v2 with about 8 GB of memory and 20 GB of disk, the full profile (event lake, entity graph, search, enrichment) is 22 services at about 12 GB, and the docs FAQ sizes a typical mid-market production load at roughly 6-8 CPU cores and 16-24 GB of RAM
- The bundled llama3.2 3B model is sized for CPU inference and the README says it is not a frontier model; stronger triage and investigation means configuring a GPU, a host model or a hosted provider, whose costs you pay
- The project states that no hosted LLM provider has been exercised in its own measurements, so its published evaluation rows for hosted models read "not measured", and it labels its benchmark numbers as substrate self-consistency measures rather than live agent accuracy
- Very young and fast-moving: the repository was created in May 2026 and shipped more than a dozen major versions (v9 to v21) between late September and November 1, 2026, and its release policy defines a major version as one that requires operator action
- The MCP server works only from a source build for now; the npm package is not published, so the documented `npx` launcher does not resolve
- Not a SIEM replacement for long-term log retention and compliance search, and it cannot see telemetry you have not connected
- Documentation is extensive but not fully consistent: some docs pages give different counts (for example of MCP tools, detection rules and memory requirements) or reference older versions than the README, which the project recounts on every change
- The docs describe an invite-only managed beta at tryaisoc.com, but on November 3, 2026 that domain redirected to the GitHub repository, so treat a hosted option as unavailable
Use cases
- Standing up a self-hosted SOC for a small team that pushes EDR, identity and cloud audit events into one pipeline and has every alert triaged before an analyst sees it
- Running AI alert triage inside an air-gapped or regulated network with a local model, so no prompt leaves the perimeter
- Handing auditors or a compliance team the Investigation Ledger or a signed evidence bundle showing exactly what the agent saw, called and concluded on a case
- Migrating Splunk, Sentinel or Elastic detections into an open-source stack and seeing which rules translate cleanly and which are refused
- Re-sweeping months of collected telemetry for sightings when a new CISA KEV entry or threat report lands
- Asking about open cases or replaying an agent's reasoning from Claude Desktop or Cursor through the MCP server
Our take
AiSOC is the only open-source, self-hostable defensive SOC platform in this directory, and an ambitious one: ingest, detection, correlation, agent triage, hunting, governed response and an MCP server in one MIT-licensed repository, with a replayable record of every prompt, tool call and verdict that you can inspect and export yourself. Its documentation is also unusually frank, labelling synthetic data, unmeasured model results and features that were fixed after not working. The catch is operational. You are adopting a multi-service platform that has shipped more than a dozen breaking major versions in about six weeks, the bundled model is a small CPU baseline, and the project has not measured triage quality with a hosted model, so verdict accuracy on your alerts is yours to establish. It makes most sense for security teams with platform engineering capacity, air-gapped or regulated environments, and MSSPs or builders who want an inspectable substrate; teams that want an AI analyst working this quarter with vendor support are better served by a commercial product. If you trial it, pin a release tag, connect one real source, and compare its verdicts against your analysts' past decisions using the built-in replay before granting any autonomy.
Who should use it
Security teams that must keep security data and prompts on their own infrastructure, regulated or air-gapped organizations that need an auditable record of what an AI agent did, MSSPs and security engineers who want an open-source base they can extend with their own connectors, detections and models, and teams with the capacity to operate Kafka, Postgres and a multi-service stack.
Who should skip it
Teams without someone to run and upgrade a self-hosted platform, organizations that need a managed service, SLAs or vendor support today, buyers expecting proven verdict accuracy without configuring and testing a stronger model, and teams looking for a long-retention SIEM.
Strengths
- Free and MIT-licensed, with the project stating there are no gated modules, Cyble-only features or telemetry
- Every agent step is recorded and replayable, and investigations export as signed evidence bundles
- Runs with no API key using a bundled local model, and supports air-gapped deployment
- Conservative response design: preview by default, triple opt-in for playbooks, approval tiers and blast-radius limits
- Unusually candid documentation that labels synthetic data, unmeasured results and what does not yet work
- MCP server and REST API make it usable from AI assistants and other tooling
Weaknesses
- Heavy self-hosted stack that you must size, secure, upgrade and monitor
- Default local model is small; serious use means configuring and paying for a stronger model
- Major versions arrive weekly or faster, each signalling required operator action
- No hosted provider measured by the project, and benchmark figures are not live accuracy
- No managed offering or commercial support is available from the project today
AiSOC pricing
Self-hosted (open source)
Free
- MIT license, no gated modules or enterprise edition
- Bundled local model, so no API key is required
- You provide the infrastructure, upgrades and any hosted model keys
Note: AiSOC has no paid plan. Running costs are your own compute and storage and, if you configure a hosted LLM provider, that provider's token charges, which the console tracks per case. The docs describe an invite-only managed beta at tryaisoc.com that is free during the beta and say a managed tier may be paid in future, but on November 3, 2026 that domain redirected to the GitHub repository. Cyble, which funds development, sells a separate closed commercial product built on the platform.
Technical specs
Available models
Where AiSOC excels
AI triage in an air-gapped network
The air-gap overlay runs a local model and blocks hosted LLM endpoints before any call is made, so alert triage and investigation work without data leaving the perimeter.
Auditable AI decisions for regulated SOCs
The Investigation Ledger and signed evidence bundles show each prompt, tool call, citation and verdict, which gives reviewers something concrete to check under SOC 2, ISO 27001 or DORA controls.
Open-source base for an MSSP or security engineering team
Multi-tenant isolation, SCIM, white-label branding, a plugin SDK and an MIT license let a team build its own service or detections on the platform without license fees.
AiSOC vs. competitors
AiSOC vs. Dropzone AI
Dropzone AI is a commercial, quote-priced AI SOC analyst delivered as a dedicated AWS-hosted tenant; it investigates alerts in place through your existing tools' APIs rather than running its own ingest and detection pipeline, and adds threat hunting, context memory and vendor support. AiSOC is free and self-hosted, ingests and detects on telemetry itself, and records every agent step, but you operate it and choose and pay for the model.
AiSOC vs. Purple AI
Purple AI is SentinelOne's agentic SOC analyst and works on data inside the SentinelOne Singularity Platform, sold as part of it. AiSOC is vendor-neutral and open source, connecting to SentinelOne among many other sources, and runs on your own infrastructure with local or hosted models.
AiSOC vs. Mave
Mave is a commercial agentic SecOps platform that queries data where it lives across existing tools, with governed response and detection tuning. AiSOC is an MIT-licensed platform you deploy yourself, with its own ingest, detection and event lake, a bundled local model and a replayable ledger of every agent decision.
Frequently asked questions
What is AiSOC?
AiSOC is an open-source, self-hostable AI security operations center. It ingests telemetry from your security tools, detects and correlates threats into incidents, has AI agents triage and investigate each one while recording every prompt and tool call in an Investigation Ledger, and proposes response actions that a human approves.
Is AiSOC free and open source?
Yes. The code is on GitHub under the MIT license and the project takes no payment from self-hosters. Your costs are the infrastructure you run it on and, if you use a hosted model instead of the bundled local one, that provider's token charges. The docs say a managed cloud offering may have a paid tier in future, but the self-hosted package stays free.
Is AiSOC a Cyble product?
AiSOC's repository says its development is funded and supported by Cyble, which pays for the engineering time and releases the work under MIT, and that the sponsorship buys no special features, gated modules or telemetry. The repository is owned by Beenu Arora, whose GitHub profile lists him as Cyble's CEO and co-founder. Cyble also ships a separate, closed commercial product built on the platform; AiSOC does not depend on it.
Does AiSOC need an OpenAI or Anthropic API key?
No. The default install bundles Ollama with a pinned llama3.2 3B model (about 2 GB) for CPU-only inference, so triage produces real verdicts without any key. Through the included LiteLLM gateway you can point each task, such as triage or deep investigation, at a hosted provider like OpenAI or Anthropic or at a self-hosted model served by vLLM, Ollama or llama.cpp.
Can AiSOC take response actions on its own?
Only if you allow it. Playbooks preview by default and run only when the deployment, the tenant and the playbook are all switched on, and each potentially impactful action maps to auto, review, escalate or reject in a per-tenant autonomy policy, with review as the default for destructive actions. Approvers need the required permission tier and cannot approve their own request, and blast-radius limits and a global advisory-only stop are available.
What do I need to run AiSOC?
The core profile needs Docker Compose v2 with about 8 GB of memory and 20 GB of free disk, plus python3 and bash; `make up` generates secrets, creates an admin account and starts 16 services. The full profile adds an event lake, entity graph, search and enrichment at about 12 GB. A Helm chart supports Kubernetes, and an air-gap overlay runs it with no outbound model calls.
Does AiSOC have an MCP server?
Yes. The AiSOC MCP server connects Claude Desktop, Cursor, Continue.dev and Cody to an AiSOC deployment so an assistant can list alerts and cases, query detections and the event lake, start an investigation, replay the agent's reasoning and preview response actions. It cannot perform actions, and it currently has to be built from source because the npm package is not yet published.
AiSOC vs Dropzone AI: what is the difference?
Dropzone AI is a commercial, cloud-hosted AI SOC analyst that investigates alerts in place through your existing SIEM, EDR and identity tools' APIs and is sold by quote. AiSOC is a free, MIT-licensed stack you host yourself that ingests and detects on telemetry, investigates with your choice of local or hosted models, and records every agent step for audit, at the cost of running the platform yourself.
Integrations & fit
Alternatives to consider
Dropzone AI
Enterprise SOCs, lean security teams that need round-the-clock alert coverage without hiring for night shifts, and MSSPs that want every alert across their SIEM, EDR, identity, cloud and email tools investigated to an evidence-backed verdict, with the team setting policy and approving containment
Custom
SentinelOne
SOC teams that run SentinelOne Singularity (or are consolidating on it, with AI SIEM for third-party data) and want alerts triaged and investigated to a verdict automatically, with response actions gated by policies they set
Custom
Mave
Enterprise SOCs, MSSPs and organizations running several SIEMs or EDRs that want AI agents to investigate every alert, hunt and tune detections across the tools they already own, with response actions under their own approval rules and no data migration
CustomAbout AiSOC
AiSOC is a full SOC stack rather than a layer on top of one. Events arrive by push to an authenticated ingest API or by pull from 91 click-and-connect connectors (Splunk, Microsoft Sentinel, Elastic, CrowdStrike, Okta, AWS and Kubernetes audit among those with vendor-specific normalization), plus collectors for CloudTrail on S3 via SQS, GCP Pub/Sub, Azure Event Hubs and syslog/CEF. A Go ingest service normalizes them onto a Kafka spine, a fusion service runs the executable part of a large Sigma-style rule library plus sliding-window rules with per-tenant tuning, and correlation groups what fires into incidents. The public agent surface is four agents (Detect, Triage, Hunt, Respond), with Triage dispatching phishing, identity, cloud and insider capabilities, and a LangGraph orchestrator for multi-step investigation. Agents call typed tools such as lake queries, graph traversals and enrichment lookups rather than writing SQL, verdicts that cite indicators the evidence never contained are demoted to human review, and the whole run is written to the Investigation Ledger and can be exported as a signed evidence bundle. Response runs through playbooks that preview by default: the deployment, tenant and playbook must all opt in before anything executes, and destructive actions default to analyst approval. The default install bundles Ollama with a small local model (llama3.2 3B, about 2 GB) so triage works with no API key, and a LiteLLM gateway lets you map each task to a hosted or self-hosted model. It also ships an MCP server so Claude Desktop, Cursor, Continue or Cody can query alerts, cases and the ledger. Development is funded by Cyble, which the project says also ships a separate, closed commercial product built on the platform. The tradeoffs: you operate a multi-service Docker or Kubernetes stack yourself, the bundled model is a CPU-sized baseline rather than a frontier model, the project says no hosted LLM provider has been exercised in its own measurements, and it is a young project that has shipped more than a dozen major versions between late September and November 2026.
Updates from AiSOC
AiSOC v21.0.0 changed correlation so alerts missing an identifier narrow their incident group instead of widening it, which the project measured as lowering its alert-reduction figure on a detection-shaped cohort from 42.6% to 24.5%. It also added end-to-end attribution of which model actually served each verdict, full-cohort recall and closure accounting in the replay scorer, a retro-hunt rehearsal mode, and asset criticality and business context feeding triage and the investigation queue.
v20.3.0 let the bring-your-own-key model selector ask a custom endpoint which models it serves, capped prompt size at the egress point, raised the auto-triage output ceiling from 512 to 2,048 tokens, and fixed several detection rules that could not fire.
Are you the founder? Claim this listing →