Operation False Copilot

AI · LINUX · difficulty 3 · ~20 min · 12 steps

A developer plugged an untrusted tool into their AI copilot. It read the repo secrets.

Briefing

A developer connected a third-party MCP server — advertised as a 'code formatter' — to their AI coding assistant on build-agent-04. Within minutes the assistant had called tools that read .env files and the git config, and something shipped those out. The model did what its connected tools let it do. A tool an agent trusts is code you have granted your data to. Find the rogue connector in the config, see what it accessed, remove it, and rotate everything it touched.

Objectives

Return to the range lobby

THREAT FEED ▲ lateral movement via wmiexec up 18% ▲ new loader family 'GHOSTPULSE' staging via DLL search-order hijack ▲ prompt injection in support queues now routine ▲ cron persistence campaign targeting exposed SSH ▲ poisoned model weights on public hubs ▲ remember: trust nothing, hash everything ▲
operator@dojo:~$