Cloudflare rolls out evidence-grounded, multi-AI-agent security operations harness for Managed Defense

Oct 8, 2026 · 4 min read

TechStaged-owned

Summary

Cloudflare introduced a built-in, multi-AI-agent security operations harness for Managed Defense that handles multiple alerts at Cloudflare scale.
The system uses OpenAI Daybreak Defense Network and Anthropic models, including GPT-5.6 Cyber and Mythos, for deeper model-backed analysis.
Initial analysis and scoring are performed with Clef, Cloudflare’s open-source decision model.

Cloudflare has introduced a built-in, multi-AI-agent security operations harness designed to handle large volumes of security alerts at Cloudflare scale. The system accelerates data gathering, detection aggregation, and accounting for missing sources as new alerts arrive, while relying on model-backed analysis for deeper context.

WHAT CHANGED FROM THE SINGLE-AGENT PROTOTYPE TO THE NEW HARNESS

The company notes that an initial, single-shot AI agent could provide useful analysis but sometimes produced results not supported by evidence. Three recurring problems were identified: context could become authoritative beyond the evidence, scope could drift beyond the correct account or time range, and lookups could time out or fail without clear "not found" signals. TechStaged has also covered Cloudflare AI Gateway Observability and Cost Controls.

HOW THE NEW HARNESS WORKS ON CLOUDFLARE MANAGED DEFENSE

Before model inference, deterministic code runs fixed reconnaissance workflows with versioned API calls to collect identity, detection history, traffic baseline, enforcement outcomes, and network observations. This reconciles inputs and makes evaluation reproducible across runs.

Front half is free of AI agents and performs data collection and boundary enforcement
Evidence packages are versioned and stored with source, version, and timestamp

SPECIALIST AI AGENTS AND THE SYNTHESIS STEP

For alerts needing deeper review, a coordinator AI agent runs four specialist AI agents in parallel: traffic analysis, customer context, global telemetry, and threat intelligence. A synthesis AI agent then combines their findings into an advisory using an approved vocabulary; the advisory can’t fetch new evidence or extend the allowed scope.

Traffic analysis checks request behavior and enforcement
Customer context references earlier alerts and analyst decisions
Global telemetry uses anonymized, aggregate signals
Threat intelligence checks existing indicators within the case

GLOBAL CONTEXT WITH PRIVACY SAFEGUARDS

The system compares alerts to patterns seen across Cloudflare’s global network but uses only aggregates to preserve customer privacy. This global context helps differentiate per-customer events from broader trends without exposing identity information.

EVIDENCE HANDLING AND VALIDATION

An evidence dossier records the alert’s track record, including previous dispositions, and the investigation cites specific items from that package. Application code on Cloudflare Workers admits evidence and validates results, ensuring that findings stay within the investigation’s scope and that invalid findings are corrected or noted as limitations.

Versioned evidence packages
Citations validated by application code
Durable Objects and state management for case context

ADVISORY GENERATION AND ANALYST WORKFLOW

An LLM-powered advisory report uses terminology familiar to Managed Defense Analysts, including affected surface, enforcement outcome, relevant controls, and next steps. Analysts can inspect evidence, revise recommendations, or group alerts into a case, while the system fixes the customer scope before any model results are presented publicly. The analyst remains responsible for final decisions and mitigations.

Evidence behind AI recommendations is visible to analysts
Analysts retain control over judgment and interventions

SCALE, BETA STATUS, AND FUTURE PLANS

The early beta is available in Cloudflare Managed Defense for eligible application-security alerts and cases. The company plans to add a Custom Managed level with more organization-specific flexibility and to explore continuous AI agents that monitor traffic for patterns beyond fixed rules and thresholds.

Beta availability for eligible customers
Future plans include Custom Managed level and continuous AI agents

WHAT THIS MEANS FOR USERS AND SECURITY TEAMS

Cloudflare’s Managed Defense customers with WAF, DDoS, Magic Transit, or related products could access the beta, potentially accelerating investigation and remediation by correlating evidence across alerts and cases. Analysts maintain final decision authority, while AI handles repetitive collection, context synthesis, and advisory generation.

SOURCES

Cloudflare Blog: Building an evidence-grounded agentic security operations harness on Cloudflare Published Oct 7, 2026 · Primary source