When individual signals are brought together, a clear picture forms. With Dynamic Risk Scoring (Q4), agent risk feeds into the same 0 to 100 risk score that Falcon Next-Gen SIEM already builds for the agent’s owner and host. When an agent, its user, and its device all climb the risk list together, the coordination behind otherwise routine activity is harder to miss.
This foundation extends to the place many of these attacks begin: the AI tooling agents depend on. AI Ecosystem Detections (Q4) add detection logic for attacks on AI services, including LLM API abuse, stolen API keys, and the compromise of coding assistants, MCP servers, and agent frameworks. When the threat starts in a trusted tool rather than the agent itself, the Falcon platform is watching there too.
Because the context around an agent doesn’t always live inside the Falcon platform, Federated Search Expansion lets teams search related data where it already lives, without re-ingestion. New connectors for GovCloud, Dynatrace, Databricks, and Snowflake are now generally available, with Splunk, Elastic, and Cribl Lake connectors following in Q4.
Act on Tough Questions
Once a risky agent surfaces, every analyst faces the same question: Did an agent’s owner do this, or did their agent do it without them knowing?
This is the work Multi-Agent Triage and Investigation (GA, announced at Fal.Con US 2026) was built for. SOC agents investigate detections end to end, with specialist agents working the case together on shared context across endpoint, identity, and the agent’s activity.
Investigation coverage already spans the full range of Falcon platform and third-party sources. In Q4, triage coverage extends to CrowdStrike Falcon® Guardian, CrowdStrike Falcon® Cloud Security, CrowdStrike Falcon® Shield, CrowdStrike Falcon® Data Security, and custom detections, including abnormal agent behavior flagged by Falcon Next-Gen SIEM. When an agent is flagged, it routes straight into an agentic investigation where the agent, its owner, and its host are already connected.
This shared context is what answers the tough question. Investigating agents can trace a session back to what the person asked for and separate it from what the agent did on its own. When the person’s request was routine and everything that followed traces to instructions they never wrote, the verdict becomes clear: The insider was the agent, not the employee.
But a verdict doesn’t stop a compromised agent that’s still running. The Unified Agentic SOAR Experience (GA, announced at Fal.Con US 2026) is where response happens. Teams build, run, and govern all automation in the SOC, from AI agents to workflows to custom apps, in plain language. They set the mission and define triggers, data, rules, and actions to decide what runs autonomously and what requires approval. A single response decision can coordinate everything that needs to follow in parallel: stopping the agent, cutting off its access, keeping it from coming back, and documenting what happened. It’s open by design and connects to any agent, model, or tool through API or MCP, so teams can govern automation across the security stack with visibility into what it’s costing and what needs attention.
See it in action:
Autonomy That Earns Its Trust
As SOCs hand more of the work to agents, the controls that keep agents accountable matter as much as the speed they deliver.
CrowdStrike builds governance into every layer of the agentic SOC. It starts by watching the SOC’s agents. Policy-Drift Agent Guardrails, powered by Falcon Guardian (Beta, Q4), screen every CrowdStrike Charlotte AI™ and AgentWorks interaction for policy drift, unauthorized tool calls, and out-of-bounds data access to catch the moment an agent steps outside its intended scope.
