Cloudflare has detailed an evidence-grounded multi-agent AI architecture operating within Cloudflare Managed Defense to streamline security alert triage at scale. The framework combines deterministic data collection with specialized language models to reduce human analyst workload while preventing AI hallucinations during incident investigations.
Addressing Single-Agent Hallucinations
Cloudflare reported that its initial prototype relied on a single general-purpose AI agent that received an entire investigation in one prompt. This configuration produced recurring issues, including scope drift across system accounts, blurred distinctions between detection hypotheses and proven exploits, and hallucinated claims unsupported by evidence. To address these limitations, Cloudflare separated evidence collection and scope boundaries from model analysis, placing initial reconnaissance into application code before model inference begins.
Deterministic Reconnaissance and Clef Triage
Before calling any AI models, deterministic application code executes a fixed set of reconnaissance workflows using versioned API calls. The system collects telemetry, customer identity, detection history, and traffic baselines, attaching timestamps and source metadata to create a replayable evidence snapshot. To filter out high-volume noise, Cloudflare uses Clef, its open-source decision model running on Workers AI, to compare incoming alerts against historic decisions. High-probability false positives bypass deeper AI analysis entirely.
Parallel Specialists and Model Integration
When an alert warrants deeper review, a coordinator AI agent assigns work to four parallel specialist agents focused on traffic analysis, customer context, global telemetry, and threat intelligence. A separate synthesis AI agent aggregates their structured findings into a single advisory without fetching new evidence or altering standardized classifications. For deeper analysis, Cloudflare incorporates models including GPT-5.6 Cyber and Mythos through partnerships with Anthropic and the OpenAI Daybreak Defense Network.
Platform Architecture and Availability
The harness runs on Cloudflare's developer platform, employing Workers for validation, Workflows for pipeline execution, D1 and R2 for storage, and Durable Objects with Flue and AI Search for state persistence. Human Managed Defense Analysts retain final responsibility for validating conclusions and applying mitigations such as WAF custom rules or rate limiting. The system is currently available in early beta for eligible application-security alerts within Cloudflare Managed Defense.
AI-generated editorial illustration; not a documentary photograph.
Source: Building an evidence-grounded agentic security operations harness on Cloudflare
SIGNAL TITANS
Leave a Reply