Cloudflare has published details of a new AI-driven system built to help its Managed Defense security analysts process the flood of alerts that modern environments generate. The company calls it an evidence-grounded, multi-agent security operations harness, and it is already running at Cloudflare scale.
The problem it addresses is what Cloudflare calls the alert paradox. A single security event can trigger a spike of related alerts across an environment, and a human analyst has to figure out which ones are connected, which are false positives, and which need immediate action. Even sophisticated Security Information and Event Management tools still leave much of this correlation work to people.
Why one AI agent was not enough
Cloudflare’s first prototype used a single general-purpose AI agent given an entire investigation at once. It produced useful analysis but also hallucinated claims that the evidence did not actually support. Telemetry, detector descriptions, policies, and threat intelligence all got flattened into one prompt, blurring their distinct roles.
Cloudflare identified three recurring failure modes: context being treated as authority (a detection is a hypothesis, not proof), scope drifting to the wrong account or time range, and failures disappearing silently when a lookup timed out instead of being logged as unchecked.
Deterministic recon before inference
To fix this, Cloudflare moved evidence collection and scope enforcement into application code that runs before any AI model analysis begins. Deterministic workflows gather identity, detection history, traffic baselines, enforcement outcomes, and network observations, each stored with its source, version, and timestamp. This fixed snapshot makes evaluations reproducible, since replaying the same data means differences in AI agent output come from interpretation rather than inconsistent retrieval.
A lightweight triage model, built on Cloudflare’s open-source Clef decision model running on Workers AI, then filters out alerts that closely match known, previously resolved patterns so they do not consume analyst or AI agent attention.
Specialist agents working in parallel
Alerts that need deeper review go to a coordinator agent that runs four specialist agents in parallel, covering traffic analysis, customer context, global telemetry, and threat intelligence. A synthesis agent then combines their findings into a single advisory, but it cannot fetch new evidence or assign a classification outside an approved vocabulary. Cloudflare says keeping each agent’s task narrow makes unsupported claims easier to catch and recommendations easier to audit.
For deeper model-backed analysis, Cloudflare uses approved OpenAI Daybreak and Anthropic models, including GPT-5.6 Cyber and Mythos, alongside its own Clef model for initial scoring.
For hosting providers and site owners who rely on Cloudflare’s security products, the approach points to faster alert triage and more consistent incident response, since the system is designed to separate verified evidence from model inference at every step.