CrowdStrike adds agent behavior detection and response to its agentic SOC
CrowdStrike has outlined new capabilities intended to help security teams detect and respond to compromised AI agents operating under trusted employee identities. Announced on November 4, 2026, the additions connect behavioral detection in Falcon Next-Gen SIEM with investigation and response functions across the Falcon platform.
The company describes an “agent insider threat” as a trusted AI agent whose behavior changes after an upstream compromise, such as tampering with a tool or introducing unauthorized instructions. Because the agent may continue using approved tools and granted permissions, individual actions can appear routine. CrowdStrike’s approach is to correlate activity across the agent, its human owner and the host it runs on, rather than assess each alert in isolation.
Behavioral signals and risk scoring
Agent Insider Threat Detection, listed for Q4, is designed to establish normal behavior patterns for AI agents and their operators. The patterns include files accessed, tools called and destinations for data. Falcon Next-Gen SIEM is intended to flag deviations even when activity does not violate an existing policy, according to CrowdStrike.
A related Q4 capability, Dynamic Risk Scoring, brings agent risk into the existing 0-to-100 risk score for the agent’s owner and host. The goal is to make coordinated changes more visible: risk signals for an agent, its user and its device can be considered together instead of treated as unrelated events.
CrowdStrike also announced AI Ecosystem Detections for Q4. These are intended to cover attacks involving AI services and supporting tools, including LLM API abuse, stolen API keys and compromises affecting coding assistants, Model Context Protocol (MCP) servers and agent frameworks. This broadens the stated detection scope beyond an agent’s own runtime activity to include components it depends on.
For context held outside Falcon, Federated Search Expansion adds search connectors without requiring data re-ingestion. Connectors for GovCloud, Dynatrace, Databricks and Snowflake are generally available. Connectors for Splunk, Elastic and Cribl Lake are planned to follow in Q4.
Investigation and coordinated response
Multi-Agent Triage and Investigation is generally available and was announced at Fal.Con US 2026. CrowdStrike says specialist SOC agents can investigate detections together using shared context across endpoint, identity and agent activity. In Q4, triage coverage is planned to extend to Falcon Guardian, Falcon Cloud Security, Falcon Shield, Falcon Data Security and custom detections, including abnormal agent behavior identified by Falcon Next-Gen SIEM.
That context is intended to help analysts distinguish an employee’s request from actions an agent took independently. Investigators can trace a session back to the person’s instruction and compare it with what the agent subsequently did. CrowdStrike says this can help determine whether suspicious behavior reflects the user’s intent or instructions the user did not provide.
The Unified Agentic SOAR Experience is generally available and was also announced at Fal.Con US 2026. It brings automation—including AI agents, workflows and custom applications—into a common environment for building, running and governing SOC actions. Teams can define triggers, data, rules and actions, and specify which operations run autonomously and which require approval. CrowdStrike says a response can coordinate actions such as stopping an agent, revoking its access, preventing it from returning and documenting the incident. The experience connects to agents, models and tools through APIs or MCP.
Controls for SOC automation
The announcement also covers safeguards for the security team’s own AI agents. Policy-Drift Agent Guardrails, powered by Falcon Guardian, are in beta and listed for Q4. They are designed to screen interactions involving CrowdStrike Charlotte AI and AgentWorks for policy drift, unauthorized tool calls and out-of-scope data access.
Human-in-the-Loop Action Requests are generally available. They pause a Falcon-native or AgentWorks agent when it invokes a sensitive tool, allowing an analyst to approve or reject the request, provide a documented reason, or inspect the agent trace before execution. Approvals for Agents and Workflows, planned for Q4, extend approval gates to an entire agent, application or workflow. Agent Interruption, generally available, allows analysts to pause or stop a SOC agent during a task.
AI Credit Visibility and Control, planned for Q4, is intended to show administrators the credits consumed by agents and workflows, let them halt consumption during a run and require approval to raise an agent’s limit. Together, these controls address both the security impact of an agent’s actions and the governance of automated SOC operations.
For enterprise teams, the announcement centers on connecting agent behavior with identity and endpoint context, then carrying that investigation into response without handing the case between disconnected tools. Several detection and governance features remain planned for Q4, while the company identifies specific investigation, response and approval controls as generally available. Organizations evaluating the capabilities should distinguish those release statuses and confirm availability against their own deployment requirements.