Opening a repository can run its code inside seven AI coding agents.
Security researchers disclosed a flaw class on 1 September that hits seven AI coding agents. The code runs before the tool asks for approval, and sometimes before you even sign in. Codex, Cursor and Goose shipped fixes, but three agents and one Claude Code path stayed open.
Why it mattersIf you let an AI agent open code you did not write, that code can run first. The weak point was not the model but the ordinary tools the agent quietly runs around it.
Read in the app — free, in 9 languages
Related stories
Walmart opened its checkout to AI agents that shop for you.2026-10-22Governments took websites offline over the AI-found PHP flaw.
2026-10-22A popular AI developer library shipped a hidden password stealer.
2026-10-22A startup says AI did most of the work on its new coding model.
2026-10-22Claude's new app checks your work accounts before you ask.
2026-10-22