Swimlane 為 Turbine 加入自訂 AI 威脅偵測及雙向 MCP 支援

An AI security analyst can investigate an alert quickly, but it cannot investigate a warning that was never generated. Swimlane is targeting that gap with new customer-defined detection capabilities in its Turbine platform, allowing security teams to apply their own logic to operational data and send the resulting alerts directly into Swimlane AI SOC.

The October 28 release also introduces two-way Model Context Protocol (MCP) support and a Hero AI Application Builder Agent. Together, the additions extend the company’s automation platform across three connected jobs: deciding which activity deserves investigation, connecting agents with tools, and building the applications that support a security workflow.

Swimlane says the capabilities are generally available. The announcement positions them as additions to existing security infrastructure, including security information and event management (SIEM), endpoint detection and response (EDR), and cloud security tools.

Giving Security Teams Control Over What Becomes an Alert

Many AI SOC products concentrate on triage, investigation and response after a detection system has identified suspicious behavior. That division becomes more difficult when organizations deploy new agents, copilots, local models and MCP servers faster than established detection products can accommodate them.

A company may need to investigate sensitive information submitted to an unauthorized AI service, unexpected use of a local model, a disabled guardrail or an agent consuming resources far beyond its intended budget. Some of those conditions are specific to the organization’s policies and environment, rather than a universal signature a vendor can define in advance.

The new Turbine capabilities let teams run their own detection logic against data available to the platform. Swimlane says a single big data action can evaluate tens of thousands of records across parallel compute nodes, then route an alert into AI SOC for investigation and governed response.

That architecture brings detection logic closer to the automation that acts on its results. Instead of separately identifying a condition and manually assembling the next steps, a team can connect its organization-specific rule to an existing investigation workflow. The usefulness of any detection still depends on whether the relevant activity is represented in the data Turbine can access.

“The security industry has spent the last two years debating how autonomously AI can work an alert, but that skips a more basic question: where did the alert come from?” said Cody Cornell, Swimlane’s co-founder and CEO.

The release also describes AI-assisted detection authoring. Teams can explain the detection they want and use Turbine’s playbook generation capabilities to help build it. Generating that logic and establishing that it reliably identifies the intended behavior remain separate tasks: organization-specific rules need to reflect the data, permissions and conditions of the environment where they run.

Investigating AI Systems as Potential Incident Sources

Swimlane describes these scenarios as AI detection and response, or AIDR: the AI system itself becomes the subject of the incident. That differs from simply using an AI model to summarize a conventional phishing or endpoint alert.

Prompt injection illustrates the distinction. An agent may encounter instructions embedded in content it is processing, creating a risk that outside material influences its behavior. A useful investigation may need to examine the agent’s inputs, tool activity and resulting actions, rather than look only for a familiar malware indicator.

Similarly, an unexpected spending surge could signal a malfunctioning or misdirected agent without resembling a traditional intrusion. Customer-defined logic gives teams a way to express what is abnormal for their own deployment, provided the necessary usage and operational signals are available.

The announced processing scale is a company specification, not a published benchmark for detecting every form of prompt injection or unauthorized AI use. The practical change is that teams gain another route for turning their own observations into actionable cases, alongside alerts supplied by established security products.

MCP Connects Agents and Automation in Both Directions

MCP provides a common interface through which AI applications can use tools. Swimlane’s October release supports both sides of that connection.

As an MCP host, Turbine allows Hero AI to connect to third-party MCP servers and use their tools during investigations and playbooks, with centralized authentication and access controls. A security workflow can therefore call an external MCP-compatible capability alongside the API integrations it already uses.

As an MCP server, Turbine exposes governed automation to external AI clients. The announcement names Claude, ChatGPT, Cursor, Gemini CLI and Microsoft Copilot as examples of clients that can invoke Turbine components, including multi-step playbooks, within existing role-based access controls.

The distinction is consequential. In the first direction, Swimlane’s agent reaches out to another system. In the second, an external assistant can initiate an approved Turbine workflow. Connecting an assistant to an established playbook gives an enterprise a way to reuse its existing automation and permissions, rather than recreate the operational process inside each AI client.

Swimlane says Turbine can combine MCP tools with conventional APIs. That makes the release relevant to environments where only some systems support MCP: teams can incorporate the protocol into a broader workflow without requiring every existing integration to change at once.

Building Security Applications With Human Review

The Hero AI Application Builder Agent addresses another source of work: configuring the record structures, layouts and workflows that a new security use case requires. Users describe what they need in natural language, and the agent creates or modifies the corresponding application.

According to the release, generated changes remain unsaved until a person reviews them. Changes are highlighted, and users can accept or reject them before they persist. That approval point matters because an application structure can affect how analysts capture evidence, move cases between stages and carry out a response.

The new agent builds on the broader Hero AI portfolio. Swimlane’s product materials describe existing agents for data ingestion, playbook generation, visualizations and investigation, as well as the ability to select models for individual agents. Application building extends that assistance to the operational workspace around those tasks.

Combining AI Reasoning With Predictable Automation

The additions feed into Swimlane AI SOC’s routing model, which directs work among deterministic automation, AI-assisted investigation and fully agentic investigation according to the reasoning required. Existing playbooks can handle familiar steps, while more complex cases can draw on agentic investigation and human review.

This is an important part of the product’s design: a workflow does not have to send every task through a model. A known response sequence can remain a conventional automation, while an agent helps interpret unfamiliar evidence or construct an investigation plan.

Customer-defined detection adds an earlier decision to that operating model: what should trigger the process in the first place? For security teams adopting AI throughout their organization, the October release offers a way to connect their own knowledge of acceptable behavior with the tools that investigate and respond when activity falls outside it.