Rene Fichtmueller
|
04349aed69
|
feat(security): v0.4.0 — three research-driven detection gaps closed
Implements hardening based on sarendis56/Jailbreak_Detection_RCS
(arXiv:2512.12069) and the Awesome-LVLM-Attack/Safety survey series.
L0 — CipherDecoder: FlipAttack, ROT13, Caesar (all 25 shifts), Morse,
Leet speak, Pig Latin, ASCII art detection with suspicion scoring.
L2 — SemanticContrastiveScanner: RCS-style harmful/benign bucket
comparison via EmbeddingStore, 20 canonical jailbreak seeds, BoW
embedding fallback for offline use.
L6 — ConversationTracker: Crescendo (+0.35), Foot-in-the-Door (+0.40),
Jigsaw Puzzle (+0.45) multi-turn escalation patterns added.
292/294 tests passing (2 pre-existing ATLASMapper failures unrelated).
|
2026-04-04 23:04:42 +02:00 |
|