// RESEARCH & LAB
Research
AI-augmented penetration testing backed by hands-on local-model work and public vulnerability research. Local LLM work: on-prem deployment and inference engines, optimization, quantization, fine-tuning, evaluation harnesses and agent/tool security applied to vulnerability research.
Some research reports are published in French.
Local LLM benchmark - code · reasoning · redteam
7 local models on a RTX 4070 Ti Super (llama.cpp, 12 fixed prompts), scored on quality + speed. GPT-OSS-20B (MoE): near champion at 4.7× the speed of dense 27B models.
- #local-llm
- #llama.cpp
- #gpu
- #evaluation
Kairos
Fine-tuned 0.3B decision model (Apache-2.0, on Laya-multilingual): zero-generation decision scoring and semantic routing with dynamic criteria; cascade deployment, ~8 ms on high-margin calls.
- #llm
- #fine-tuning
- #decision-routing
- #apache-2.0
CVE-2026-86950 — Apple CoreGraphics Integer Overflow
Technical study of an integer-overflow issue affecting Apple iOS and macOS. The public advisory classifies the CVE impact as an out-of-bounds write.
- #CVE-2026-86950
- #research
GitHarvest3r
Recon tooling for source and secret discovery, with custom Nuclei templates.
- #recon
- #Nuclei
Nerv0usR4bbit
Open-source offensive security tool.
- #tooling
Ping-0f-Death
Open-source security research project.
- #research
Soft-Script-V.2
Open-source security tooling.
- #tooling
Microsoft Office Word MSHTML RCE (2021) — Metasploit Ruby exploit
Ruby exploit module for Metasploit.
- #ruby
- #metasploit
CrackEverything6 — contribution
Open-source contribution (2023).
- #open source
Malicious WordPress plugin — contribution
Open-source contribution (2024).
- #open source