skip to content

In Cloudflare's WAF, what does the attack score in cf.waf.score measure, and how should a team start acting on it?

level: middleimportance: should knowfreq 30%

answer

  1. machine learning, not signatures
  2. low number means likely attack
  3. four classes from attack to clean
  4. a detection needs a rule to act

basics

~20 s

WAF attack score is a machine-learning detection that gives each request a score from 1 (almost certainly malicious) to 99 (likely clean), catching modified payloads that miss signatures. It acts only through a rule, so start with a strict threshold such as 20.

solid answer

~40 s

Attack score complements the managed rulesets. Signatures catch known payloads, and the model also scores variants an attacker has altered to slip past them. `cf.waf.score` runs from 1 to 99, where **lower means more likely malicious**. There are per-vector scores in `cf.waf.score.sqli`, `cf.waf.score.xss` and `cf.waf.score.rce`, and 100 marks a request the system chose not to score. The full fields need Enterprise. Business plans get only `cf.waf.score.class` (`attack`, `likely_attack`, `likely_clean`, `clean`). It is a **detection**: it fills fields and blocks nothing until a custom rule or rate limiting rule uses it. Cloudflare's suggested start is a rule on `cf.waf.score le 20`, watched for a few days (with Log on Enterprise), then moved to Managed Challenge or Block. Blocking everything below 50 catches legitimate requests.

code

plaintext · 1 line
plaintext
cf.waf.score le 20

go deeper

for a junior

Recall the scale's direction, 1 likely malicious and 99 likely clean, and that the score only acts through a rule.

for a middle

Explain the four classes and their ranges, the per-vector fields, the 100 and 0 special values, and which plan gets which field.

for a senior

Show a staged rollout from a strict threshold to enforcement, and how a scoped attack score rule backs up a managed-rules exception.

for a principal

Judge where machine-learning scoring replaces signature tuning effort, and where a business case for the Enterprise fields holds up.

## What the score is Cloudflare's **WAF attack score** is a traffic **detection**. A machine-learning model looks at each request and estimates how likely it is to be an attack. Its job is to cover a gap in signature matching: - The **Cloudflare Managed Ruleset** matches known attack signatures and has a very low false-positive rate. - Attackers change known payloads (re-encoding, fuzzing, reordering) so they no longer match a signature exactly. - The attack score model rates those variants too. Cloudflare recommends running **both** managed rules and attack score. The system automatically detects and decodes Base64, JavaScript Unicode escape sequences and URL-encoded content anywhere in the URL, headers and body. ## Reading the numbers The direction surprises people who know the OWASP Core Ruleset, where a higher threat score is worse. Here **a low score means a likely attack**: | Value | Meaning | | --- | --- | | `1`–`20` | class `attack` | | `21`–`50` | class `likely_attack` | | `51`–`80` | class `likely_clean` | | `81`–`99` | class `clean` | | `100` | reached the system but was not scored (shown as *Unscored*; not usable as a class in expressions) | | `0` | Logpush only: the request never reached the scoring stage, for example because an earlier rule mitigated it | The fields are: - `cf.waf.score`: the global score, derived from the per-vector scores but **not** their sum. - `cf.waf.score.sqli`, `cf.waf.score.xss`, `cf.waf.score.rce`: SQL injection, cross-site scripting and command injection or remote code execution. - `cf.waf.score.class`: the class string from the table above. ## Plans - **Enterprise**: all numeric fields and the class. - **Business**: only `cf.waf.score.class`. - **Free and Pro**: no attack score. ## Why it blocks nothing by itself Like other detections, attack score only **populates fields**, visible as filters in Security Analytics. Nothing happens to traffic until a **custom rule** or a **rate limiting rule** uses a field in its expression. The managed rulesets are different: they come with actions. The attack score needs a rule you write. Writing and ordering those rules is covered in the lesson on custom rules. ## A safe first rule Cloudflare's own rollout guidance runs in three steps: 1. **Create a strict rule.** On Enterprise, `cf.waf.score le 20`. On Business, `cf.waf.score.class eq "attack"`. 2. **Monitor it** for the first few days in Security Events, checking that the threshold fits your traffic. Enterprise customers can start with the *Log* action, which Cloudflare's actions reference lists as Enterprise-only. 3. **Tighten the action** to *Managed Challenge* or *Block* once matches look right. Cloudflare advises against blocking every request below 50 on the score alone, because the `likely_attack` band (21–50) includes legitimate requests. To catch more of that band, combine a higher threshold with other conditions, such as a sensitive URI path, or a bot score where Bot Management is available. ## Where it fits beside the rulesets - On an endpoint where a managed-rules exception skips noisy rules, for example a rich-text save, a strict attack score rule can still act on requests the model rates as clear attacks, so the carve-out does not leave the path unwatched. Check first that the endpoint's legitimate requests stay out of that band. - Attack score and **bot score** answer different questions. Attack score asks whether the request content looks like an attack. Bot score asks whether a person or an automated client sent it. The bot material covers the latter. - A request mitigated in an earlier phase may never be scored, so read a `0` in logs as "not reached", not as "malicious". ## Looking before writing the rule Because detections populate fields whether or not a rule uses them, you can study the score on real traffic before you enforce anything: - In **Security Analytics**, which shows all incoming requests and not only the ones a security feature acted on, filter by the attack score class or a score range and see which paths and clients land in the `attack` band. - Compare that band with what the managed rulesets already mitigate. Requests in the `attack` class that no managed rule matched are the extra coverage the score buys you. - Look for legitimate patterns that score low, such as admin forms posting markup or code snippets, before a rule starts challenging them. - From the filtered view, the dashboard offers **Create custom security rule**, which turns the current filters into a rule expression you can review before you deploy it. This look-first habit is what keeps a strict threshold from turning into a support queue on launch day.

  • Logpush records show cf.waf.score values of 0 on some requests. Were those attacks?
    No. In Logpush a `0` means the request never reached the attack score stage, for example because a custom rule or DDoS protection had already mitigated it. It does not appear in the dashboard. A `100` is different again: the request reached the system and was deliberately left unscored.
  • How does attack score help on a path where a managed-rules exception skips some OWASP rules?
    The exception stops the skipped rules from scoring that endpoint's legitimate markup, but the attack score is still computed for the request. A strict custom rule such as `cf.waf.score le 20`, scoped to the same path, keeps acting on requests the model rates as clear attacks, once you have checked that the endpoint's legitimate saves do not score that low.

  • A high cf.waf.score means the request is probably an attack.
  • The global attack score is the sum of the SQLi, XSS and RCE scores.
  • Turning on attack score detection blocks attacks automatically.
  • Blocking every request scoring below 50 is the recommended starting rule.
  • Business plans can write rules on the numeric cf.waf.score field.