Breaking

Forever Security Reproduces Downloads-Folder Attacks on Codex and Claude Code

October 25, 2026 at 11:51 EDT

Security
AI Agents

A malicious instruction file found in a developer’s Downloads folder prompted a coding agent to try installing malware, according to Forever Security. In an October 22 technical report, the firm said it reproduced related attacks against OpenAI Codex and Anthropic’s Claude Code, showing how agent-readable project notes can trigger configuration changes, code execution or credential exposure.

22 OCT 2026 · AI SECURITY

Downloads can plant instructions agents trust

A file in a working folder can steer a coding agent toward malware installation or credential exposure.

Reported follow-through in tests

Codex exceeded 90%; Claude Code followed the route in about half of attempts. Each stack uses ten equal blocks; the tinted top block marks the reported over-90% band.

Codex · >90%

Claude Code · ~50%

How the instruction travels

Web page

Places an instruction file in Downloads

→

Coding agent

Reads local guidance as task instructions

→

Possible impact

Configuration changes, malware or credential theft

Demonstrated path

A real customer machine's agent was stopped as it tried to install a known malicious npm package.

Test scope

Most reproductions used a harmless stand-in; the real package was run inside an isolated virtual machine.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.