Crate wafrift_wafmodel 

wafrift-wafmodel — the WAF decompiler

Stop searching a black box. Reconstruct the WAF’s decision boundary as an executable symbolic automaton, then turn evasion from search into deduction:

• P1 — Decompile. Active-learn the WAF (the learn module) over a WafOracle into an Sfa, spending the minimum membership-query budget. Emit it as a provenance-stamped artifact.
• P1 — Mine. Intersect the learned pass-language with an attack grammar offline to harvest minimal-edit bypasses with no further live queries.
• P2 — Solve. Compose the learned WAF view with the pipeline’s normalization transducers and solve for inputs that survive every stage (the double-decode trick, rediscovered — not hard-coded).
• P3 — Dominate. The same model drives constrained adversarial evasion of ML-WAFs and provable hole-closure for defenders.

Everything here is zero-config and pure-Rust: no GPU, no external Coraza, no network required for the core. Acceleration (vyre/GPU, live HTTP oracles) is strictly additive.

The crate is built bottom-up; each module is landed complete (no stubs) before the next depends on it. This file only declares modules that are fully implemented.

Re-exports

pub use artifact::LearnedModel;

pub use artifact::Provenance;

pub use booster::WafBoosterScorer;

pub use canon::CanonView;

pub use canon::Channel;

pub use canon::Segment;

pub use canon::canonicalize;

pub use ensemble_dilution::RuleGroup;

pub use equiv_bridge::norm_mismatch_members;

pub use equiv_bridge::sink_for_tag;

pub use equiv_bridge::solution_member;

pub use equiv_query::ChainedEq;

pub use equiv_query::PacBound;

pub use equiv_query::SampledEq;

pub use equiv_query::UcbBanditEq;

pub use equiv_query::WMethodEq;

pub use error::Result;

pub use error::WafModelError;

pub use filter_profile::DecodeGap;

pub use filter_profile::FilterProfile;

pub use filter_profile::TokenFinding;

pub use filter_profile::TokenProbe;

pub use filter_profile::Verdict;

pub use filter_profile::battery_from_toml;

pub use filter_profile::characterize;

pub use filter_profile::default_battery as default_filter_battery;

pub use filter_profile::probe_decode_gaps;

pub use harden::ClosureReport;

pub use harden::synthesize_closure;

pub use learn::Alphabet;

pub use learn::BoundedExhaustiveEq;

pub use learn::EquivalenceOracle;

pub use learn::LearnReport;

pub use learn::kv_learn;

pub use learn::l_star;

pub use learn::l_star_budgeted;

pub use learn::passive_learn;

pub use mine::attack_grammar;

pub use mine::mine_bypasses;

pub use mine::minimal_bypass;

pub use mine::waf_diff;

pub use mlwaf::MlEvasion;

pub use mlwaf::MlWaf;

pub use mlwaf::evade_ml;

pub use normalize::Transform;

pub use normalize::apply_chain;

pub use oracle::ChannelSet;

pub use oracle::FnOracle;

pub use oracle::Rule;

pub use oracle::SimRegexWaf;

pub use oracle::WafOracle;

pub use origin_probe::FnReflector;

pub use origin_probe::OriginScan;

pub use origin_probe::ReflectionOracle;

pub use origin_probe::detect_origin_normalization;

pub use origin_probe::scan_origin;

pub use outcome::Outcome;

pub use sfa::BytePred;

pub use sfa::Sfa;

pub use sfa::StateId;

pub use solve::Solution;

pub use solve::solve_bypass;

pub use transduce::Pipeline;

pub use transduce::Stage;

pub use transduce::json_unescape;

pub use transduce::url_decode_once;

Modules

artifact

The learned-model artifact: a decompiled WAF, serialized.

booster

WAFBooster importance scoring (paper: “WAFBooster: Automatic Boosting of

canon

Canonicalize a wafrift_types::Request into the ordered set of byte-segments a CRS-class WAF actually inspects.

ensemble_dilution

#101 Multi-sub-score ensemble dilution.

equiv_bridge

Zero-downstream-change bridge: solver/preimage output expressed as the canonical EquivPayload the rest of the ecosystem already consumes.

equiv_query

Query-economical equivalence oracles — the strategies that make decompiling a live WAF affordable.

error

Error type for the WAF-decompilation engine.

filter_profile

Filter characterization — learn WHICH attack tokens a live WAF actually policies, by differential probing.

harden

The defensive dual: from a decompiled WAF’s holes, synthesize the minimal rules that close them, prove zero new false positives against a benign corpus, and prove the class is closed (no attack-grammar member survives the hardened config).

learn

Active automaton learning over the WAF oracle.

mine

Offline bypass mining over the decompiled model.

mlwaf

Constrained black-box evasion of ML-WAFs.

normalize

CRS input transformations — the decoding a ModSecurity/Coraza-class WAF applies to a variable before matching a rule against it (t:urlDecodeUni, t:htmlEntityDecode, t:lowercase, …).

oracle

The oracle the learner queries: does this request reach the app?

origin_probe

Origin-normalization fingerprinting — measure which decode/normalize stages a target’s origin applies, so the P2 solver TARGETS its preimage to the real pipeline instead of speculatively trying every canonical sink.

outcome

The two-valued projection of a WAF verdict that the learner reasons over.

sfa

Symbolic finite automata over the byte domain.

solve

Composition + preimage solver — the part that turns “encoding tricks” from hand-written rules into emergent solutions.

transduce

Pipeline-stage transducers.

Functions

default_crs_ruleset

The shipped Tier-B OWASP-CRS-derived ruleset (XSS 941 + SQLi 942), embedded so wafrift audit/harden work zero-config with no files to fetch. Parse with SimRegexWaf::from_toml.