·
13 commits
to main
since this release
v4.4.1
This tag was signed with the committer’s verified signature.
eebe132
This commit was signed with the committer’s verified signature.
v11b Neural Threat Detection Model
- Model: Char CNN-BiLSTM v11b, 1.6M params, Latin-1 vocab (256), 256-char input
- Evasion Resistance: 99.8/100 (up from 97/100 at v10)
- In-scope misses: 0 (all 41 in-scope ATLAS techniques detected)
- Phase3c k6 validation: 100% TPR (52/52 blocked), 0% FPR (0/38 flagged)
What's New
- All 50 augmentor transforms synced to match evasion suite
- keyWalkReverse text normalization (QWERTY right-shift inverse, homoglyphs, l33t, zero-width Unicode)
- 2,600-test adversarial evasion suite (52 ATLAS payloads × 50 transforms)
- OPSEC hardening: pre-commit hooks activated, CODEOWNERS, PR templates
- Version refs synced: VERSION file, docker-compose image, README badges
Model Hash
SHA-256: 8e13c793c32816aa0f6e2af13ffadd4f38f707b4ac8906b56ddfa77da51ea8e5
Docker
docker pull ghcr.io/aegisgatesecurity/aegisgate-platform:4.4.0
``