Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

An X-Ops analysis of how AI coding agents trigger endpoint detection rules designed for adversaries AI coding agents (Claude Code, Cursor, Codex, and others built on skill packs such as GStack) are showing up in customer environments. They write code, install dependencies, automate browser tasks, and troubleshoot failures by trying alternative approaches.

At Corelight Labs, our mission is to help organizations stay a step ahead of evolving threats. When our researchers came across Censys' detailed write-up on ScoutC2, a rapidly growing open-source command-and-control (C2) framework favored by threat actors, we knew we needed to bolster community defenses quickly.

It was time to write another book. That’s what I thought when I heard that Corelight wanted to update its 2021 book on network detection and response (NDR). Tamara Crawford, who owned the project, scheduled a meeting with me and asked if I might be interested in helping, depending on who might write the text.

"Exploit mitigation" stalled around 2008. The attacks didn't. AI turns a patched bug into a working exploit in hours. Why endpoint exploit mitigation stalled in 2008, and what a default-on mitigation layer looks like now. The core of my last post was an asymmetry. An attack can unfold in two steps or 20, but the vocabulary it draws from never grows.

Your SOC probably already has detections. The problem is that many of them don't behave like a managed security capability. They behave like a pile of alerts. Analysts close noisy rules because they have to protect their queue. Engineers keep adding logic because coverage gaps are real. Leaders ask whether the program is improving, and the usual answers are weak. Alert counts go up. Tuning tickets pile up.

Counter Threat Unit (CTU) researchers investigated two interconnected threat groups known as Vect and TeamPCP. The two groups announced a formal operational partnership in late March 2026 to combine TeamPCP’s credential harvesting and data theft capabilities with Vect’s ransomware deployment infrastructure in a widespread campaign involving supply chain attacks and the extortion of multiple organizations.

For years, the dominant threat against remote identity verification was the presentation attack: someone holding a printed photo up to a camera, wearing a mask, or playing a pre-recorded video on a phone screen. The industry responded with increasingly sophisticated anti-spoofing technology and vision-based detection models, and the standards to test their effectiveness followed. But many of today’s most sophisticated fraudsters don’t bother with the camera at all.

The first time an MCP (Model Context Protocol) server felt real to me, it wasn't because of a clean demo. It was because of the noise. TL;DR: The harness matters more than the protocol, and the evidence matters more than both. MCP earns its keep when it shortens the path from a good security question to trustworthy evidence, and almost everything interesting about making that work happens in the harness wrapped around the model. In this series, I will cover how to build an MCP for an AI SOC.