Cloudflare puts evidence collection before AI security analysis

Managed Defense combines coded evidence gathering, specialist agents and a human analyst’s final decision.

The entrance and reception desk at Cloudflare’s San Francisco office.
Archive photograph, 17 December 2021: Cloudflare’s office at 101 Townsend Street, San Francisco. Company context for its security-agent announcement. Resized and converted to WebP; full frame retained. HaeB · CC BY-SA 4.0

Cloudflare has described a security investigation system that collects evidence through deterministic code before asking models to interpret it. Clef handles initial triage; specialist agents examine traffic, customer context, global telemetry and threat intelligence. A synthesis agent combines their findings for a human analyst.

The company says the design addresses failures seen in a general-purpose prototype, including unsupported conclusions and scope drift. Missing evidence remains distinct from a negative finding. Access is an early beta for eligible Managed Defense customers. VCT’s takeaway: keeping retrieval and authorization outside model inference makes an agent’s recommendations easier to audit.

Source

[Cloudflare Blog announcement](https://blog.cloudflare.com/agentic-security-operations/).

Cloudflare Blog