A fake Claude Code extension in the Cursor marketplace tried to backdoor developer machines over SSH

The story that matters most this morning is a supply chain hit on the people building AI software. CSO Online reports that a malicious version of Anthropic's Claude Code extension was distributed through the Cursor marketplace and tried to install a persistent SSH backdoor on developers' machines. Cursor has pulled it. Developer laptops hold source code, cloud credentials and model API keys all in one place, so a backdoor there is about as valuable as access gets.

What to do: if anyone on your team installed a Claude Code extension in Cursor recently, check ~/.ssh/authorized_keys and any new SSH keys or services on those machines today, and rotate credentials that lived on any box you can't vouch for. My take is that editor extension marketplaces have become the new npm, and most shops still treat them as personal preference rather than software they need to approve.

Chainalysis has disclosed a data breach that SecurityWeek says came from attackers using offensive agentic techniques to get into its systems and internal tools. Coming right after the ARTEX campaign CrowdStrike pinned on China-backed Vapor Panda last week, this is no longer a research curiosity. What to do: look at your internal tools the way an automated attacker would, because an agent will happily try every exposed admin panel and stale token you forgot about, and it won't get bored.

Sophos, via Help Net Security, describes the Aur0ra ransomware group using AI agents to move through victim networks, harvest credentials and climb to domain admin. What to do: alert on fast, high-volume credential use and lateral movement from a single host, since speed is the giveaway when the operator is a script that doesn't sleep. The practitioner's point is that the playbook hasn't changed much. What changed is how cheaply a mediocre crew can run it.

Two more worth your coffee. Schneier passes along a comment that roughly 3% of all critical software bugs ever recorded were found in the last month, which tells you patch queues are about to get longer, so budget for it. And Wired found around 400 apps and sites that pose as ordinary AI tools while hiding features for generating abusive deepfake images and video, which is a useful reminder to check what AI apps your users actually have installed.