WAF test report

Results at a glance

Version 0.32.4, tested on 2026-11-02 against the public chaitin/blazehttp corpus: 33 877 raw HTTP messages (658 malicious, 33 219 benign).

Metric	Result
Effective-calibre interception	96.6% (616/638)
Full-calibre interception	93.8% (617/658)
Benign traffic affected	0.06% (19/33 219)
Accuracy	99.82%
Static-tier p95	≤ 2 ms

Which number should you use? The 96.6% denominator is 638: 20 of the 658 malicious samples are excluded because no proxy-side WAF can decide them — 19 POST/PUT requests with no Content-Length / Transfer-Encoding (an empty request body under RFC 7230) and 1 sample whose payload hides in a URL fragment, which the HTTP parser strips before building the request line. The full-calibre figure (93.8%) is kept for reference.

Comparison with other products

Product / mode	Samples	Malicious	Detection rate	Benign	False positives	Accuracy	Time
PingWAF Strict+body	33 877	658	93.8% (617)	33 219	0.06% (19)	99.82%	p95 2 ms (loopback)
PingWAF Normal	33 877	658	60.2% (396)	33 219	0.01% (4)	99.21%	p95 2 ms (loopback)
SafeLine free, strict	33 669	575	76.17% (438)	33 094	0.22% (73)	99.38%	64.34 ms (public)
SafeLine free, balanced	33 669	575	71.65% (412)	33 094	0.07% (23)	99.45%	70.05 ms (public)
ModSecurity PL4	33 669	575	94.61% (544)	33 094	52.46% (17 362)	48.34%	28.89 ms (public)
ModSecurity PL1	33 669	575	69.74% (401)	33 094	17.58% (5 819)	82.20%	31.15 ms (public)
Cloudflare free	33 669	570	10.70% (61)	32 780	0.07% (23)	98.40%	288.96 ms (public)

The PingWAF rows are self-measured (effective calibre 96.6%, see the section above); the remaining figures are quoted as-is from the blazehttp README and were not re-tested. The sample sets differ (33 877 vs 33 669 samples; 658 vs 570–575 malicious), and third-party timings are public-network round trips while PingWAF figures are loopback verdict latency — compare detection and false positives, not milliseconds.

Detection by category

Category	n	Strict+body	Normal
SQL injection	20	85.0%	85.0%
XSS	30	96.7%	46.7%
Path traversal	25	96.0%	80.0%
Command injection	12	83.3%	25.0%
SSRF	13	100%	7.7%
Deserialization	6	100%	16.7%
CRLF injection	5	100%	60.0%
SSTI	5	100%	60.0%
XXE	2	50.0%	0%
Log4Shell	1	100%	100%
Scanners	1	100%	0%
Encoding & obfuscation	538	93.9%	61.9%

What changed in this round

1. Chinese search and telemetry values containing and 1=1 are no longer judged as SQLi (7 samples), and .github paths are no longer treated as dotfile probes (1 sample) — the four tier counts are unchanged (396/473/500/617) while benign samples affected drop from 12/20/12/27 to 4/12/4/19.
2. End-to-end coverage of compressed attack bodies: all 8 malicious gzip payloads blocked, both benign ones passed (2/2).
3. Challenges (503) and blocks (403) are now counted separately.

Tier	Interception	blocked	challenged	Benign affected	p95
Normal	60.2% (396/658)	396	0	0.01% (4)	2 ms
Normal+body	71.9% (473)	473	0	0.04% (12)	2 ms
Strict	76.0% (500)	490	10	0.01% (4)	2 ms
Strict+body	93.8% (617)	607	10	0.06% (19)	2 ms

Where the numbers come from

A release build of the static TOML data plane (pingwaf -c with a mock origin); the 33 877 raw messages are replayed byte for byte over a raw HTTP/1.1 socket on loopback, 16 concurrent connections with 5 ms spacing; only the 403 block page and the 503 challenge page count as interceptions — protocol-layer and method-level rejections never count; 250 engine unit tests pass; and every figure is reproducible with the commands below.

Reproducing

git clone --depth 1 https://github.com/chaitin/blazehttp /tmp/blazehttp
export BLAZEHTTP_ROOT=/tmp/blazehttp
cargo build --release --bin pingwaf --features=full
./tools/waf-bench/run_suite.sh --conf tools/waf-bench/conf/level-strict-body.toml --suite full --group strict-body --out results/0.32.4/strict-body --concurrency 16 --delay-ms 5