AI Pulse · Under the Radar

A red-team testing skill for Claude called Black-cat picked up 145 GitHub stars in two days, almost entirely from Chinese AI-security communities, with no English-language coverage yet. The repo is bilingual and frames itself around hypothesis-evidence-driven offensive testing. What makes it notable is not the tool itself but the format: it’s packaged as a “skill” for Claude Code, not a standalone app, which means it slots into the agent’s workflow rather than sitting outside it.

This matters because skills are starting to behave like distribution units. Instead of building a separate tool that calls an API, developers are writing small, reusable capabilities that agents can invoke directly. Black-cat spreads faster because anyone running Claude Code can drop it in and start testing. The barrier to adoption is lower, and the skill travels with the agent, not as a bolt-on. We’re seeing the same pattern in other corners of the ecosystem, but this one moved quickly in a niche that doesn’t usually share early.

What it means for operators

If skills become the default way to extend agent capability, you’ll need a way to evaluate, version, and govern them inside your own systems. A red-team skill that works today might behave unpredictably tomorrow if the underlying model shifts, and you won’t know until something breaks in production. This is exactly the kind of thing we build into an AI command centre: a layer that tracks what skills your agents are using, logs their behaviour, and flags drift before it becomes a problem. The speed at which Black-cat spread, documented in the original repo, suggests skills will move faster than traditional software, and governance needs to keep pace.