v1.1.3 — MLSecOps Practical Reference Guide
KV Cache security section for LLM inference (reconstruction threats, PromptPeek side channels, CAG, KV-Cloak-class controls) plus maintainer site link on the docs site.
| Format | Link |
|---|---|
| Source (ZIP) | https://github.com/MHaghighian/MLSecOps/archive/refs/tags/v1.1.3.zip |
| Markdown | chapters-en/ in this repository |
| Site | https://mhaghighian.github.io/MLSecOps/ |
| Maintainer | https://mhsec.me |
| Word (optional) | python scripts/build-docx.py --render-mermaid |
Highlights
- Ch.7 — KV Cache security — what KV Cache is (Transformer mechanism, not a product); direct reconstruction (Inversion / Collision / Injection) vs cross-tenant prefix timing (PromptPeek);
CAGdurable-cache risks; Mature isolation baseline + Emerging KV-Cloak-class obfuscation - Cross-links: multi-tenant sections, Appendix E.1.3, Ch.16 GPU isolation; bibliography entries for Luo et al. (NDSS 2026), Wu et al. (NDSS 2025), Chan et al. (2024)
- Docs site footer / home: mhsec.me
Haghighian, M. (2026). MLSecOps Practical Reference Guide (v1.1.3).
Zenodo. https://doi.org/10.5281/zenodo.21206781
Full notes: RELEASE_NOTES.md · CHANGELOG.md