mirror of
https://github.com/samvallad33/vestige.git
synced 2026-07-24 23:41:01 +02:00
b107fd5d7f
Swarm-audited (DeepSeek V4-Pro) + Claude-verified gaps in the Microglial
Firewall's screening normalization:
- fold_confusable: add Greek homoglyphs that are TRUE visual Latin matches
(ο→o ι→i α→a ε→e ρ→p τ→t χ→x κ→k β→b). Upsilon/nu deliberately omitted
(not reliable homoglyphs). Math-alphanumerics left as a TODO (need per-block
offset tables to fold correctly).
- is_format_or_space: add U+00AD soft hyphen, U+034F, U+2061..U+2064 so they
can't hide inside a keyword and split tokenization (system\u{AD}: bypass).
- has_external_destination: broaden exfil prepositions (via/through/using/over/onto).
- has_long_base64_run: accept the base64url alphabet (-, _).
- strip ordered-list gutter (1. / 2)) before the role-prefix check.
The benign-marker/quoted-trigger suppression (a deliberate false-positive
tradeoff) is intentionally left unchanged. clippy clean; 30 firewall tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
||
|---|---|---|
| .. | ||
| benches | ||
| src | ||
| Cargo.toml | ||