Cloudflare Introduces Beta Multi-Agent Security Harness for Managed Defense

Cloudflare has introduced a multi-AI-agent security operations harness within Cloudflare Managed Defense to address the "alert paradox," where simultaneous alerts overwhelm human analysts. The system automates evidence gathering and context aggregation, allowing analysts to focus on resolution rather than data collection. It leverages models from the OpenAI Daybreak Defense Network and Anthropic partnership, including GPT-5.6 Cyber and Mythos, for deep analysis, while using Clef, Cloudflare’s open-source decision model, for initial scoring and triage.

Architecture: Recon First, Inference Second

The harness avoids the pitfalls of single-agent systems, such as hallucinations and scope drift, by separating deterministic data collection from AI inference. Before any model analysis begins, application code executes fixed reconnaissance workflows to collect customer identity, detection history, and traffic baselines. This creates a versioned, reproducible snapshot of evidence, ensuring that differences in AI findings stem from interpretation rather than inconsistent data retrieval.

Triage and Specialist Agents

To reduce noise, Clef running on Workers AI quickly filters alerts likely to be false positives, skipping them from deeper analysis. For alerts requiring review, a coordinator agent runs four specialist agents in parallel:

• Traffic analysis: Reviews request behavior and enforcement outcomes.
• Customer context: Checks historical alerts and analyst dispositions.
• Global telemetry: Compares activity with privacy-preserving, Internet-wide signals from Cloudflare’s network.
• Threat intelligence: Validates indicators against admitted case data.

A synthesis agent combines these findings into an advisory report, restricted to an approved vocabulary and unable to fetch new evidence or cross tenant boundaries. The system explicitly distinguishes between "not checked," "checked with no match," and "checked with evidence of absence" to handle incomplete data transparently.

Human-in-the-Loop Workflow

The final advisory provides recommended next steps, such as rate limiting rules or WAF custom rules, but Managed Defense Analysts retain full responsibility for decisions and mitigations. The underlying infrastructure uses Cloudflare Workers, Workflows, D1, R2, and Durable Objects to manage state and validation. The beta is currently available for eligible application-security alerts, with plans to expand to a Custom Managed level and continuous monitoring agents in future quarters.