Cloudflare launches evidence-grounded multi-agent AI harness for Managed Defense alert triage (early beta)
Cloudflare has described a multi-agent AI security operations harness in its Cloudflare Managed Defense service, now in early beta for eligible application-security alerts and cases. It is meant to reduce analyst overload from correlated alerts. Deterministic code collects evidence and enforces scope before any model runs, which addresses hallucination, scope drift and silent lookup failures seen in a first single-agent prototype. The harness uses Clef, Cloudflare's open-source decision model, for triage and scoring. For deeper analysis it uses approved OpenAI Daybreak and Anthropic models, including GPT-5.6 Cyber and Mythos. This is product news with no direct threat to defenders.
Technical Details
- Deterministic reconnaissance workflows with versioned API calls collect customer identity, detection history, traffic baseline, enforcement outcome and network observations. Each item is stored with its source, version and timestamp, so the snapshot can be replayed.
- Clef, running on Workers AI, scores each alert against its recon data. Likely false positives skip specialist analysis, and known high-volume noise is classified as passive.
- A coordinator agent runs four specialists in parallel: traffic analysis, customer context, global telemetry (aggregates only, with no other customer's data) and threat intelligence. A synthesis agent combines their typed findings into one advisory. It cannot fetch new evidence or choose a classification outside the approved vocabulary.
- A versioned evidence package records scope, time anchor, admitted evidence, policy versions and coverage gaps. Application code validates that every specialist citation exists, belongs to the investigation and supports the claim.
- The advisory distinguishes "checked, no match" from "checked, evidence supports absence" and makes no classification when evidence is insufficient.
- The platform stack is Workers, Workflows, D1, R2, Durable Objects, Flue and AI Search.
Remediation
Not applicable, as this is not a vulnerability or incident. Suggested mitigations in the advisories (rate limiting rules, WAF custom rules, DDoS changes) are applied by Managed Defense Analysts for fully managed customers. Other customers receive recommendations in the dashboard. Analysts remain responsible for final decisions.
Linha do tempo
-
Cloudflare launches evidence-grounded multi-agent AI harness for Managed Defense alert triage (early beta)
ler a análise desta fonte ocultar
Cloudflare has described a multi-agent AI security operations harness in its Cloudflare Managed Defense service, now in early beta for eligible application-security alerts and cases. It is meant to reduce analyst overload from correlated alerts. Deterministic code collects evidence and enforces scope before any model runs, which addresses hallucination, scope drift and silent lookup failures seen in a first single-agent prototype. The harness uses Clef, Cloudflare's open-source decision model, for triage and scoring. For deeper analysis it uses approved OpenAI Daybreak and Anthropic models, including GPT-5.6 Cyber and Mythos. This is product news with no direct threat to defenders.
Technical Details
- Deterministic reconnaissance workflows with versioned API calls collect customer identity, detection history, traffic baseline, enforcement outcome and network observations. Each item is stored with its source, version and timestamp, so the snapshot can be replayed.
- Clef, running on Workers AI, scores each alert against its recon data. Likely false positives skip specialist analysis, and known high-volume noise is classified as passive.
- A coordinator agent runs four specialists in parallel: traffic analysis, customer context, global telemetry (aggregates only, with no other customer's data) and threat intelligence. A synthesis agent combines their typed findings into one advisory. It cannot fetch new evidence or choose a classification outside the approved vocabulary.
- A versioned evidence package records scope, time anchor, admitted evidence, policy versions and coverage gaps. Application code validates that every specialist citation exists, belongs to the investigation and supports the claim.
- The advisory distinguishes "checked, no match" from "checked, evidence supports absence" and makes no classification when evidence is insufficient.
- The platform stack is Workers, Workflows, D1, R2, Durable Objects, Flue and AI Search.
Remediation
Not applicable, as this is not a vulnerability or incident. Suggested mitigations in the advisories (rate limiting rules, WAF custom rules, DDoS changes) are applied by Managed Defense Analysts for fully managed customers. Other customers receive recommendations in the dashboard. Analysts remain responsible for final decisions.