Phishing has grown in volume and sophistication, with attackers moving beyond generic email blasts to techniques such as two-factor man-in-the-middle (MitM) attacks, QR codes that evade detection rules, and AI-generated personalized messages at scale. Anti-Phishing Working Group (APWG) data shows incidents climbing year over year.
Cloudflare has reported a similar rise across 2023 and 2024. Abuse reports arrive from anyone who spots potentially abusive behavior on sites using its services, and its Trust & Safety investigators and engineers respond to those complaints. Before the automation work, that response was manual: investigators reviewed each complaint, the reporter's allegations, and the reported site's content to judge whether it was phishing or malware. With a growing customer base and rising phishing volume, the process became unsustainable.
From manual triage to an automated pipeline
The results came in two stages. In the first half of 2024, Cloudflare resolved 37% of phishing reports through automated means, with a median time to action on hosted phishing reports of 3.4 days. In the second half, after deploying the new tooling, automated systems resolved 78% of reports and median time to action on hosted phishing reports dropped to under an hour.
A group of internal abuse experts built the system on the Cloudflare Developer Platform, which removed most of the scaling concerns and let the team focus on the detection engine itself. The data feeding it comes from several places: network-wide insights, the internal Email Security product, external feeds from trusted sources, and years of accumulated abuse report processing data.
How a URL is assessed
- Each submitted URL first goes to a detection Worker, which runs it through the Cloudflare URL Scanner. The scan returns rendered HTML, network requests, and site attributes.
- The HTML and page resources are then passed to in-house machine learning classifiers for reputational information.
- Indicators of compromise (IOCs) are sent to threat feeds and domain categorization tools to flag known malicious sites and existing categorizations.
- A set of rules and heuristics, derived from how investigators have historically handled similar reports and from observed bad-behavior patterns, decides whether the URL is phishing and what remediation to apply.
Reporters themselves can be wrong or may deliberately try to weaponize the abuse process, so the rules have to account for false positives against legitimate sites. Such reports can erode customer trust and create incidents, so the automation includes steps to disregard erroneous ones.
Platform components
Several primitives on the Developer Platform hold the pipeline together: KV stores report summaries at scale, and Durable Objects keep running counters for an unlimited set of tracked attributes. Hyperdrive provides access to the internal Postgres database holding abuse reports, Queues manages scanning jobs, Workers AI runs the machine learning classifiers, and D1 stores detection logs for evaluating efficacy. A Remix Pages UI ties everything together, presenting the engine's analysis to investigators for follow-on work and for reviewing inconclusive results.

Intelligence that feeds outward
The same data used to speed up abuse processing also improves customer-facing products. Analyzing abusive patterns — common phrases in phishing attempts, attacker infrastructure, coordinated campaigns across multiple domains — strengthens application security, email security, and threat intelligence offerings. Brand Protection customers can report suspected abuse directly from the Cloudflare dashboard, and Trust & Safety can act on similar threats network-wide, including for customers who are not Brand Protection users. Cloudflare also partners with customers and outside experts to learn about trends seen in their own mitigation efforts, which informs tightly targeted rules for emerging phishing patterns.
For anyone who spots phishing passing through Cloudflare's network, the company maintains an abuse reporting form and, for technical users who want to report programmatically, an abuse reporting API.
What customers can do
- Enterprise customers can speak with their Customer Success Manager about enabling Brand Protection, which is included by default for all enterprise customers.
- Existing Brand Protection users should update their brand's assets so legitimate sites and logos can be distinguished from possible phishing activity.
- All customers should confirm their abuse contact is current in the Cloudflare dashboard.



