Abstract. Single-agent AI code review has a structural flaw: it presents hallucinated findings with the same confidence as real ones, and it never learns from being wrong. Gossipcat is an MCP server for Claude Code that attacks both problems with a single mechanism. Several agents review a change independently, then cross-review each other’s findings against the source code; the verified outcomes become grounded reward signals that reshape how future work is routed and how each agent is prompted. No model weights are updated — the learned policy is a set of markdown files. This post explains why that design works, the engineering problem that nearly sank it, and what the system’s own development history reveals about its limits.

Root cause analysis and PoC for a Microsoft SQL Server Stack Overflow Vulnerability by reversing svl.dll.

This is the write-up of the OneTwoSeven machine from HackTheBox.

Greetings! With solving Fortune machine, I finished half of the number of machines on HackTheBox. At present, Fortune has not retired yet. But I decided to write it’s writeup. I will share this blog post when the machine is retired. So, if you are reading this blog post right now, it means you are looking into the past.