netgoat-ai/WAF-Bench

Dataset Summary

A small, hand-curated benchmark of 77 labeled HTTP requests for evaluating Web Application Firewalls (WAFs) and ML-based request classifiers. Each record is a complete, replayable request labeled malicious or benign, with the expected WAF action, an OWASP Top 10 (2021) mapping, a difficulty rating, and the evasion technique used (if any). Benign records include deliberate false-positive traps (e.g. SQL keywords in prose, apostrophes in names).

Intended Uses

• Measuring detection rate and false-positive rate of WAFs (ModSecurity/OWASP CRS, cloud WAFs, custom rules).
• Regression-testing rule changes and paranoia-level tuning.
• Quick sanity-check evaluation of ML classifiers for malicious request detection.

Out-of-Scope Uses

• Not a substitute for a full security assessment or penetration test.
• Not suitable as sole training data: the set is small and synthetic.
• Payloads are detection strings, not working exploits; they should not be used to attack systems you do not own or have permission to test.

Dataset Structure

File: test.jsonl (one JSON object per line; loaded as the test split).

Field	Type	Description
id	string	Unique ID (waf-001 ...)
category	string	Attack class or benign
label	string	malicious or benign
expected_action	string	block (malicious) or allow (benign)
method	string	HTTP method
path	string	Request path
query	object	Query parameters (values are as sent, pre-encoding where the case is not an encoding test)
headers	object	Request headers
body	string or null	Raw request body
difficulty	string	easy, medium, or hard
evasion	string or null	Obfuscation technique, if used
owasp	string or null	OWASP Top 10 2021 mapping (malicious only)
notes	string	Free-text annotation

Example

{"id": "waf-001", "category": "sqli", "label": "malicious", "expected_action": "block", "method": "GET", "path": "/products", "query": {"id": "1' OR '1'='1"}, "headers": {"Host": "shop.example.test", "User-Agent": "Mozilla/5.0"}, "body": null, "difficulty": "easy", "evasion": null, "owasp": "A03:2021-Injection", "notes": "classic tautology"}

Statistics

• Total: 77 (malicious: 57, benign: 20)
• Records using an evasion technique: 14

Category	Count
benign	20
sqli	11
xss	9
cmd_injection	6
ssrf	6
path_traversal	5
ssti	3
sensitive_file	3
lfi	2
xxe	2
jndi	2
nosqli	2
scanner	2
rfi	1
crlf	1
open_redirect	1
http_smuggling	1

Difficulty	Malicious	Benign
easy	27	13
medium	20	2
hard	10	5

Creation

Authored manually from widely documented attack patterns (OWASP Top 10, public WAF test corpora conventions). All hostnames use the reserved .test TLD or example.com. No real user data, credentials, or live targets are included.

Recommended Evaluation

Replay each request against the WAF under test (in a lab environment) and compare the outcome to expected_action. Report:

• Detection rate (recall) on malicious records, overall and per category
• False-positive rate on benign records
• Results broken down by difficulty and evasion
• Precision/F1 if comparing classifiers

Note that "block" should be defined up front (HTTP 403, anomaly score above threshold, etc.), and the WAF configuration (rule set version, paranoia level) should be reported alongside results.

Limitations and Biases

• Small and uneven: 77 records; several categories have only 1-3 examples, so per-category rates are noisy.
• Class imbalance: malicious records outnumber benign roughly 57:20; false-positive estimates have wide confidence intervals.
• Single-request scope: no multi-step attacks, session state, rate-based or bot-behavior detection.
• Synthetic traffic: does not reflect the distribution of real production traffic.
• Borderline labels: a few benign items (e.g. a query about writing a <script> tag) are intentionally ambiguous; some teams may reasonably label them differently.
• Coverage gaps: no GraphQL, WebSocket, deserialization, or file-upload payloads; HTTP smuggling is represented by one case.
• Payload styles reflect common public techniques and may be outdated relative to current attacker tradecraft.

Ethical Considerations

Contents are limited to non-weaponized detection strings for defensive testing. Run replays only against systems you own or are authorized to test.