Security · AI · Engineering
Notes from real-world security research, AI agents, and software engineering.
$ whoami
security researcher / builder
$ ls topics/
AI Security Agent Security Vulnerability Research Code Auditing Offensive Engineering
$ wc -l archive
147 posts
01 / LATEST
The newest research and engineering note
Replacing an LLM with a 307M encoder for tagging personal data in API traffic, at 0.93 macro-AP. Covers data splits, sliding windows, DAPT, long context vs. windows, a GPU memory fix, and the label-suppression problem I only found right before release.
02 / MORE
Vulnerability research, agent security, and engineering practice
A retrospective on the development of Zbot, an LLM-driven web vulnerability scanner, covering its architectural evolution, sub-agent design, tooling decisions, benchmark results, lessons learned, and unresolved challenges.
This article describes integrating SQLMap into an MCP-based AI workflow to validate SQL injection vulnerabilities efficiently, preserve complete authenticated HTTP requests, and handle long-running scans asynchronously.
What is LOSS? LOSS is a differentiable mathematical function that quantifies the error between a model’s predictions and the ground truth. A larger value indicates a larger error. Note that loss is not necessarily related to the actual effectiveness of model fine-tuning; it is only used to observe whether the fine-tuning process is proceeding normally. During fine-tuning, it serves as the optimization objective, guides parameter updates, evaluates training performance, and helps determine when to stop training.
An overview of the key hyperparameters for model fine-tuning, including learning rate, epochs, batch size, LoRA rank, and validation set size, with recommended values and practical considerations.
Set up a Python environment, develop and test an MCP service locally, and build and publish the package to PyPI.
An introduction to SecGPT-Distill, an experimental open-source model that uses fine-tuning and knowledge distillation to overcome existing models’ restrictions when handling security-related questions.
Fine-tuning and RAG can often achieve similar results, so the right choice depends on practical requirements such as data freshness, latency, explainability, general-purpose capabilities, and cost.
This is an important question because modern large language models are already highly capable and may not require fine-tuning; goals can often be achieved through RAG or prompt engineering, depending on the scenario.
This article explains why human preference alignment was omitted when training a small security-focused model and compares traditional RLHF with DPO in terms of workflow, complexity, data requirements, resource demands, and performance.