I am Nathan (m00dy), a security researcher and penetration tester focused on AI/ML runtime security, binary format and parser verification, weight-space analysis, and adversarial tradecraft simulation.

My work centers on finding memory-safety, type-confusion, and deserialization defects across the native C, C++, and Rust software stack that loads and executes machine learning models and binary containers, alongside building open-source tooling for red-team infrastructure, post-exploitation research, and defensive telemetry validation.

• Notebook & Writing: m00dy.sh
• The Unchecked Atlas: unchecked.sh
• GitHub: github.com/professor-moody

• Secant (Docs) - Rust and Python toolkit for Weight-Space Programming (WSP) covering zero-evaluation spectral weight auditing (secant inspect / read), multi-subspace Gram-Schmidt composition with closed-form Target-Shift Calibration (secant compose), zero-copy mmap and Q8_0 block-aligned surgery (secant patch / quant patch), and subspace-locked LoRA micro-training (ProjectThenHealLinear).
• Crucible (Docs) - Structure-aware grammar and mutation framework for ML model formats (GGUF, ONNX, SafeTensors, TFLite, FlatBuffers, Protobuf) that pairs schema-driven mutators with sanitizer-instrumented harnesses and stack-hash crash deduplication.
• The Unchecked Atlas - Interactive vulnerability research atlas mapping parser boundaries, container deserialization hazards, and native runtime memory-safety invariants across modern systems and ML stacks.

• Scorch (47★) & ludus_scorch (19★) - Simulated MS SCORCH Env (Scorch) and its automated Ludus deployment role for repeatable red/blue cyber exercises.
• aipostex (6★) & aipostex-lab - AI/ML post-exploitation tradecraft framework and companion lab environment for researching security boundaries across deployed inference services and model pipelines.
• ludus_litterbox_role (20★) - Ansible role for deploying the LitterBox malware analysis and payload testing sandbox into Ludus virtual environments.
• Hemlock - Modular command-and-control (C2) framework built to study extensible implant communication protocols and operator workflows.
• halo-forge (9★) - Benchmark and reinforcement-learning training framework (Halo-Forge) for AMD Strix Halo systems.

• ML Runtimes, Compilers & Tokenizers:
  • tensorflow/tensorflow (16 merged PRs across core/framework, core/kernels, core/grappler, core/util/tensor_bundle, and core/lib/jpeg/webp: #129806, #129690, #129689, #129688, #129671, #129663, #129658, #129657, #129654, #129627, #129626, #129620, #129613, #129516, #129508, #129495)
  • openxla/xla (2 merged PRs via Google Copybara across xla/tsl/lib/io: #51992/#52109, #52416/#52459)
  • microsoft/onnxruntime (#33517, plus maintainer-merged fix #28751)
  • onnx/onnx (#8037)
  • iree-org/iree (#25088)
  • google/sentencepiece (#1368)
  • EricLBuehler/mistral.rs (#2495, #2282)
  • ollama/ollama (#17062)
  • ModelCloud/GPTQModel (#2950)
  • leejet/stable-diffusion.cpp (#1750)
  • sonos/tract (#2425)
  • turboderp-org/exllamav3 (#310)
  • mozilla-ai/llamafile (#1039)
• Vector Indexes & Columnar Storage:
  • lance-format/lance (#8138)
  • unum-cloud/USearch (#776)
• Computer Vision, Image Codecs, Text Shaping & OCR:
  • opencv/opencv (#30577, #30262, #30261) & opencv/opencv_extra (#1431, #1430)
  • harfbuzz/harfbuzz (OpenType text shaping: #6563, #6557, #6556, plus maintainer-merged fixes #6546 and #6542 with author credit)
  • tesseract-ocr/tesseract (OCR: #4671, #4670)
  • libjxl/libjxl (JPEG XL codec: #5075)

• CVE-2025-66455 (InternLM/lmdeploy) - Untrusted deserialization in model loading (GHSA-2vh9-42vm-xmv2).
• CVE-2026-14647 (onnx/onnx) - Heap out-of-bounds read in ONNX shape inference (#8037).
• CVE-2026-70659 (leejet/stable-diffusion.cpp) - Heap out-of-bounds write in the imatrix loader (GHSA-6q7q-2v9c-x8m7, fixed in #1750).
• CVE-2026-78147, CVE-2026-78148, CVE-2026-86317, CVE-2026-17500, CVE-2026-17501 & CVE-2026-18581 (ggml-org/llama.cpp) - ggml-rpc and GGUF parser/loader vulnerabilities.
• CVE-2026-10298, CVE-2026-17512 & CVE-2026-17513 (ggml-org/whisper.cpp) - Heap out-of-bounds read, null dereference, and reachable assertions in model/audio parsing.
• CVE-2026-84286 / CERT/CC VU#369611 (turboderp-org/exllamav3) - Out-of-bounds function-pointer dispatch in reconstruct kernel (#310).
• CVE-2026-86288 (ModelCloud/GPTQModel) - Device out-of-bounds read from unchecked g_idx in Triton dequantization (#2950).
• CVE-2026-86289 (ollama/ollama) - Integer-underflow panic in GGUF v1 string parsing (#17062).
• CVE-2026-75090 (EricLBuehler/mistral.rs) & CVE-2026-75093 (sonos/tract) - GGUF token-ID out-of-bounds panic (#2282) and NNEF parser crash.
• CERT/CC VU#681158 (nmslib/hnswlib & spotify/voyager) - Heap out-of-bounds write during HNSW index deserialization when header capacity relationships are violated.