> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Cloudflare set frontier AI models on its own firewall and turned 49 findings into new rules
- URL: https://globalfeed.ai/en/cloudflare-set-frontier-ai-models-on-its-own-firewall-and-turned-49-findings-into-new-rules/
- Published: 2026-09-29T19:55:05.000Z
- Updated: 2026-09-29T19:55:05.000Z
- Description: Cloudflare cast the latest AI models as attackers and tested its own firewall across 45 scenarios. Human review left 49 findings, which led to new SSRF rules. The company also made Threat Signals, which reads open threat reports with AI, free for every account.
- Author: GlobalFeed Editor
- Tags: Cloudflare, WAF, ai security, cybersecurity, x-Cloudflare, dil-en, elle

**Cloudflare** described on September 29, 2026 how it tested its web application firewall (WAF) with the latest AI models. According to the company, customers kept asking the same question: is your WAF ready for frontier AI models? Cloudflare tried to answer it by attacking its own system.

## How the test worked

The model played the attacker with no access to source code or firewall rules, seeing only part of each response. It started from known attacks the firewall already blocks, then changed the encoding, the place of the payload in the request or the way the target address was written to try to get through. For each attempt, one call proposed a new variation and a second call reviewed the result. The model did not send requests itself; separate software handled the allowlist, the attempt limit and logging. The target was an authorized customer's staging environment.

A total of 1,107 attempts were recorded across 45 scenarios. XSS, SQL injection, path traversal and Log4j attacks were almost fully covered. After human review, 49 findings relevant to the firewall remained, and 48 of them were command injection or SSRF, which forces a server to make requests to internal addresses. One example was the cloud metadata address written with a trailing dot: decimal and octal forms had been blocked, but this form got a redirect instead of a block. Cloudflare stresses that no successful response came from the server and no data leaked.

The findings led to two new SSRF rules and an updated one, rolled out automatically to all customers in July and August. The company did not say which AI models it used. Two of its lessons stand out: adding more attempts to the same scenario stopped helping after a point, as the model began repeating earlier ideas. And without human review, not a single real finding emerged.

## Threat Signals is free for everyone

**Threat Signals**, announced the same day, follows open threat reports that the user picks through RSS. AI skills summarize each report, extract its indicators of compromise (IoCs) and tag them using the account's existing tags. Each indicator becomes a record in the account's private threat data and can be turned into a firewall rule. Starting today, every Cloudflare account gets one RSS feed and 30 days of storage for free, and Cloudforce One's Threat Events platform is also open to everyone at no cost. More feeds and custom skills are part of paid plans.

The tool used in the WAF test is not offered to customers; Cloudflare says it is building a similar service, Adaptive Security, that will regularly test selected URLs with AI agents, but gives no date. The test was done in a single customer environment with a black-box method, and results from a test with an attacker that knows the rules will be published separately. To see how the approach holds up elsewhere, it is worth following the work of independent security researchers.

Sources: [Cloudflare: We tested our own WAF with frontier AI models](https://blog.cloudflare.com/adaptive-ai-waf-testing/?ref=globalfeed.ai), [Cloudflare: Introducing Threat Signals](https://blog.cloudflare.com/threat-signals/?ref=globalfeed.ai), [Cloudflare: Adaptive application security for the AI era](https://blog.cloudflare.com/ai-era-framework/?ref=globalfeed.ai).