Notes on the AppSec pipeline: building AI security systems, human-led pentesting, hardening findings into fixes, and watching attack surfaces.
The Hugging Face incident showed agents that recognized an attack was out of scope — and joined it anyway. A policy the model can reason around is a suggestion, not an authorization boundary.
October's coding-agent vulnerability wave barely touched the models. GitSpawn, Plugin4Shell, trojanized hooks, and explosive prompts all hit the harness — and what actually stops them.
A viral post claims Qwen3.8-27B does 381 tokens per second on a single RTX 3090. The number is real — and it's not chat speed. It's speculative decoding doing exactly what security agent workloads need.
Calibrated decision models — not chat models — are the right tool for agent control flow. Latency, cost, and determinism math, and how deciders fit security agent pipelines.
Open deciders, MIT frontier weights, and a desktop Blackwell box: an entire agentic security pipeline run locally, privately, at fixed cost.
Zhipu's 320B MoE flagship ships MIT-licensed weights that fit a 128GB desktop box. What frontier-adjacent open weights mean for private security workloads.
The AppSec pipeline: build, break, harden, watch — one loop, enter anywhere, every stage feeding the next.
Continuous recon and attack-surface monitoring as the always-on stage: what it watches, what it catches, and why point-in-time testing isn't enough.
A pentest report that sits in a ticket queue changed nothing. Turning findings into code fixes, guardrails, detectors, and automation — the Harden stage.
Human-led testing of web, APIs, networks, and LLM systems — how to tell a real assessment from a checkbox scan, and what you should actually get.
What an AI security engineering engagement actually looks like: agent pipelines, eval harnesses, detection automation — and what a 1–4 week build delivers.
The workflow behind 3-day pentests: where our custom tooling saves hours, and where a human still makes every call.
Missing input validation between the user and the model is the most common and most dangerous gap. How to close it.
Traditional pentesting and AI-system security are different disciplines. Shipping an AI product means you need both.