Urgent.News

What's breaking now, across thousands of outlets.

AI

A web page can drop AGENTS.md or CLAUDE.md into ~/Downloads. Codex obeyed in over 90% of tries, Claude Code in about half

According to Forever Security, a web page can save a file into ~/Downloads without a click, and a coding agent started in that folder reads the instruction file it finds there. On 2026-10-22 Forever Security reported catching a planted AGENTS.md doing exactly that on a customer's Mac. Putting their write-up next to the Codex and Claude Code documentation shows the attack has to land in a…

A web page can drop an instruction file into the Downloads folder, and an AI coding agent launched from that location will read and follow the instructions. On October 22, 2026, Forever Security observed a customer's Mac downloading an AGENTS.md file and a coding agent reading it, attempting to install a malicious npm package. Although the file was harmless Markdown, it was treated as a guide for the agent to execute.

Testing revealed that Codex followed the AGENTS.md in over 90% of attempts, while Claude Code was more resistant, with a similar CLAUDE.md successfully manipulating Claude in about half of cases. Both coding agents require the folder to be trusted before executing the instructions. Codex loads project-specific config files when trusted, but AGENTS.md cannot override system-level settings like base URLs. In contrast, Claude Code allows AGENTS.md to set ANTHROPIC_BASE_URL when trusted.

To protect against these attacks, users should check which files agents read and which settings can be altered by folder contents. For Codex, examine ~/.codex and project directories, while for Claude Code, focus on CLAUDE.md and CLAUDE.local.md files. Users may also want to review the trust settings and permissions prompts to ensure no unauthorized files can be executed.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Read the original at dev.to →

More in AI

948 Linux Root Exploits in One Run: What OpenAI's Kernel Pipeline Actually Proves — and What It Doesn't

OpenAI's November 4 research post, "A thousand fewer paths to root," reports a striking result: a prototype pipeline powered by an upcoming cyber model scanned the Linux kernel in mid-October 2026 and…

  • OpenAI published research on 948 Linux kernel LPE exploits with patches.
  • Pipeline involved five stages with human review at each step.
  • 948 unique LPEs were verified as local privilege escalations requiring low-privileged access.

More from Thursday 5 November →