Building an evidence-grounded agentic security operations harness on Cloudflare

Overview of Cloudflare's Agentic Security Operations Harness

Cloudflare's Managed Defense system employs a multi-agent AI architecture to address the challenges of analyzing security alerts at scale. This system separates the deterministic collection of evidence from model-based inference, ensuring that recommendations are grounded in verifiable data. The approach involves a series of specialized AI agents that work in parallel to analyze different aspects of an alert, such as traffic behavior, customer context, global telemetry, and threat intelligence. This structured workflow reduces the risk of hallucinations and scope drift, which are common pitfalls when relying on a single general-purpose AI agent.

Key Components of the System

The system begins with a deterministic reconnaissance phase that collects essential data before any model-based analysis occurs. This phase includes gathering customer identity, detection history, traffic baselines, enforcement outcomes, and network observations. Each piece of data is stored with its source, version, and timestamp, ensuring reproducibility and traceability. This initial step allows the investigation to connect the behavior that triggered an alert with the control that fired and its outcome.

Following reconnaissance, a lightweight triage model filters out noise by comparing each alert with its reconnaissance data. This step helps identify alerts that are likely false positives and skips them from further analysis by specialist AI agents. The triage model uses Clef, an open-source decision model, to quickly assess the likelihood of an alert being a false positive based on historical data and previous analyst decisions.

Specialist AI Agents and Their Roles

For alerts that require deeper investigation, a coordinator AI agent runs four specialist AI agents in parallel. These agents focus on specific tasks: traffic analysis, customer context, global telemetry, and threat intelligence. Each agent examines its assigned domain and provides findings that are later synthesized into a single advisory by a synthesis AI agent. This approach ensures that each task remains narrow and focused, making it easier to catch unsupported claims and audit recommendations.

The global telemetry specialist agent compares an alert with patterns observed across Cloudflare's global network. This agent works with aggregated data to preserve customer privacy, ensuring that no individual customer's data is shared with other customers. This global context is combined with customer-specific data to provide a more comprehensive view of potential threats.

Handling Incomplete Evidence and Decision-Making

The system is designed to handle incomplete evidence by distinguishing between three states: not checked, checked with no matching result, and checked with evidence supporting absence. When evidence is insufficient, the system avoids making classifications or disposition recommendations. Instead, it relies on Clef to score the evidence and determine whether it is sufficient for a decision. This ensures that recommendations are based on the available evidence and do not overstep the data's limitations.

Finally, an LLM-powered agent produces an advisory report using terms that Managed Defense Analysts are already familiar with, such as affected surface, enforcement outcome, relevant controls, and next steps. This report is reviewed by analysts, who can inspect the evidence, investigate further, revise recommendations, or group alerts into a case. The system ensures that the model never receives authority to act on behalf of analysts or cross tenant boundaries.

Source: Cloudflare. Analysis based on publicly available information; no hands-on testing is claimed.