wafrift · 2026-07-30 · product

Meet wafrift: a programmable WAF-evasion engine.

A web application firewall (WAF) you have never attacked is a ruleset you are trusting on faith. wafrift attacks it for you. It mutates real attack payloads through encoding layers, grammar rules, and protocol tricks, fires them at your WAF, and runs evolutionary search over what gets through. The bypasses that win are persisted to a per-WAF gene bank, so the next scan starts from everything the last one learned.

Install it

cargo

cargo install wafrift-cli

from source

git clone https://github.com/santhreal/wafrift && cd wafrift && cargo build --release

wafrift is for infrastructure you own or are authorized to assess. Point it at your own WAF, never at someone else's.

Mutate a payload without a target

wafrift evade is the offline half. Pipe a payload in, get bypass variants out, no target required. Every strategy is addressable as a path, so you can include or exclude exactly the layers you want:

echo "' OR 1=1--" | wafrift evade --only encoding/url/triple
echo "' OR 1=1--" | wafrift evade --exclude grammar/sql/comments

Encodings cover URL, double and triple URL, Unicode, hex, base64, and chained combinations. Grammars are dialect-aware: a SQL mutation understands tautologies and comment syntax, an XSS mutation understands event handlers and tag structure, so the variants are attacks, not byte noise.

Scan a live WAF and get a verdict

wafrift scan fires variants at a live target and classifies each response with a multi-signal oracle: block, bypass, challenge, or rate-limit. It honors Retry-After instead of tripping your own rate limits. The JSON report ends in a blunt verdict:

wafrift scan https://your-app.example/search?q=test

"waf_bypass": {
  "waf_in_play":      true,
  "bypass_confirmed": true,
  "verdict":          "bypass_confirmed"
}

Exit codes match the verdict. 0 means a bypass is confirmed. 4 means the WAF fought back and no variant won. 6 means no WAF is in play on that parameter at all, which is a different finding: your measurement was never about the WAF. CI can gate on each outcome separately.

Know which WAF you are fighting

wafrift detect fingerprints the WAF, CDN, and origin on four independent axes, so stripping one layer does not hide the rest: HTTP response headers and body (over 160 vendor rules), the DNS CNAME chain (37 vendor rules), reverse DNS on the leaf IP, and a BGP origin-ASN lookup. eBay's custom proxy banner resolves to Akamai through its CNAME chain. Stripe's Server: nginx resolves to AWS through the ASN axis. The fingerprint decides which mutation families are worth spending requests on.

The part that learns

Static mutation lists hit the same wall as static rulesets: the WAF vendor updates, the list does not. wafrift treats bypass discovery as a search problem. Hill-climbing, simulated annealing, tabu search, novelty search, and MAP-Elites (an illumination algorithm that keeps a map of diverse elite candidates rather than one best answer) all run over the mutation space. Every confirmed bypass is persisted to a per-WAF gene bank under ~/.wafrift, so the next engagement against the same stack starts with zero discovery cost.

When you already have a working bypass, wafrift distill shrinks it. It applies Zeller's ddmin algorithm (delta debugging: systematically remove chunks and keep whatever still fails the test) to find the smallest payload that both bypasses the WAF and still executes as an attack. A semantic oracle gates every reduction, so the minimal form cannot collapse into a benign string that merely passes the filter.

Drop it into an existing workflow

As a proxy. wafrift-proxy sits between Burp, Caido, or mitmproxy and the target, applies evasion at the upstream forward, and records bypasses to the gene bank. MITM mode with TLS impersonation (Chrome, Firefox, Safari ClientHellos, with matching header order) is built in.
As a one-shot report. wafrift oneshot runs detect, fingerprint, and bypass-probe against one target and stitches the results into a single Markdown writeup.
As a replay. wafrift replay re-fires a known-good bypass against any target, exits 0 on bypass, 2 on block. Regression tests for a WAF rule change become one line of CI.
As libraries. Encoding, grammar, detection, smuggling, evolution, oracle, and strategy are separate crates on crates.io. You can embed the mutation engine without the CLI.

Honest status

wafrift is beta. Local stacks (ModSecurity paranoia levels 1 through 4, Coraza, BunkerWeb, naxsi) are exercised in CI under wafrift-bench/. Cloud-WAF coverage (Cloudflare, AWS WAF, Akamai, Imperva, F5) is still sparse, so treat those results as preliminary and tell us what you hit. The full version history is in the changelog.

Try it

github.com/santhreal/wafrift - MIT OR Apache-2.0, no telemetry. Install instructions and the feature overview live at santh.dev/wafrift. Use it against assets you own or are authorized to test.