Research, publications & open source

Portable security review skill for Claude Code, Codex, Cursor, Copilot, and Windsurf. Packages OWASP LLM 2025, GenAI, and agentic security guidance into instructions that coding agents can apply during code generation, edits, and PR review. Covers 33 categories across LLM, RAG, MCP, tool, memory, permission, approval-flow, and orchestration risks.

Research MCP server that gives MCP-compatible agents access to web search, webpages, PDFs, GitHub repositories, Hacker News, Stack Overflow, OpenAlex, Semantic Scholar, arXiv, and YouTube transcripts without requiring API keys. Supports stdio for desktop clients and HTTP for agent stacks that connect to a local MCP endpoint.

Authorization Type Attestation protocol for agentic communication systems, now published as an active IETF Internet-Draft. TLS secures the channel; ATA identifies the class of party that authorized the session. The draft targets H2H, H2AI, AI2H, and AI2AI communication states across TLS, QUIC, MLS, WebRTC, MCP, and A2A-style systems without replacing identity layers such as mTLS, OAuth, SPIFFE, or FIDO2.

English repo and PDF research artifact for the June 2026 paper "Resolution Capacity: Beyond Reconstruction MSE for Progressive Approximate Nearest-Neighbor Search". Shows why reconstruction MSE is the wrong objective for progressive RAG retrieval, introduces Resolution Capacity, compares Hadamard and PCA rotations, includes Python reference experiments, Rust-backed 25k/1M validation, SAQ comparison notes, and R6 beam+rerank evidence.

Cross-Provider Adversarial Review framework where independent AI reviewers from different providers review a document blindly across iterations until consensus convergence. The design composes model strengths through the document as shared medium, uses web-grounded validation each round, and reduces herding, single-model blind spots, and default complimentary bias.

Benchmark and live demo for vector compression on real GloVe-200 embeddings. The app normalizes vectors, applies randomized Hadamard rotation, fits Lloyd-Max scalar codebooks for 1, 2, 4, 6, and 8 bits, rebuilds HNSW, and plots Recall@10 against compression. The practical result: 8-bit is the safe setting, 6-bit is aggressive but usable, and 4-bit needs reranking.

Recursive Convergent Inference proposes dynamic module expansion for neural inference: start from a minimal active expert subset, expand through an affinity graph, and stop when the model's next-token distribution converges. Evaluated on OLMoE-1B-7B across 150 reasoning tasks, with the notable finding that the model's computational demand diverges from human difficulty labels.