This is less a one-off than a preview: as AI coding assistants become the default developer surface, their extension ecosystems are turning into the same attack vector that plagued npm and VS Code. The attackers' branding is telling — piggybacking on the Claude Code name is the 2025 equivalent of a fake crypto wallet. Cursor deserves credit for the blunt, fast response of killing the entire marketplace, but the incident exposes a real gap: there is no mature vetting pipeline for AI tool extensions yet, and up to 5,000 installs in 11 hours shows how quickly trust-based distribution spreads malware. Expect copycats until marketplaces introduce mandatory review and sandboxing.

MarkTechPost is covering a practical but modest research release, and the headline deserves some skepticism — aggregators love quoting the top-end 30.9% when the typical band is closer to 18%, and SWE-Bench numbers say little about messy production codebases. The underlying idea is sound, though: agentic coding burns huge token budgets on redundant context and tool output, so a learnable filter that leaves the agent untouched is a genuine cost lever. In our view the open question is generalization beyond benchmark replays. It also signals where the fight is drifting — as raw model prices fall, the savings now come from squeezing the agent stack itself, and plugins like this are that efficiency layer.