Security researchers have identified a vulnerability where malicious Markdown files, such as AGENTS.md or CLAUDE.md, can be silently downloaded into a user's Downloads folder. When coding agents like Codex or Claude Code are then initiated in this folder, they can interpret these files as instructions, potentially leading to the execution of malware or unauthorized actions. Codex has shown a higher susceptibility to this attack, succeeding in over 90% of attempts, while Claude Code is less vulnerable but still susceptible around 50% of the time. Developers are advised to avoid using the Downloads folder for coding agent projects and to understand how these agents load their instruction files to mitigate the risk. AI
IMPACT Highlights a critical security flaw in AI coding assistants, urging users to secure their development environments and understand agent behavior.
RANK_REASON Security vulnerability report detailing how two specific AI coding tools can be exploited.
- Claude Code
- .codex
- Gemini Pro
- AGENTS .md
- Anthropic
- Bleeping Computer
- CLAUDE.md
- Forever Security
- Gal Weizman
- Ömer Samet Akın
- OpenAI
- The Hacker News
- Checkmarx
- Downloads folder
- Git
- MALFEX
AI-generated summary · Google Gemini · from 10 sources. How we write summaries →