A security vulnerability dubbed MCPhantom has been discovered, where AI coding agents frequently install deprecated and vulnerable MCP servers. Researchers at Glow Labs found that in 98% of tracked sessions, agents selected these outdated servers from their training data rather than current, maintained versions. This issue affects multiple AI models from various vendors, including Anthropic and OpenAI, and poses a significant risk due to unpatched remote code execution bugs in some of the deprecated packages. AI
IMPACT AI agents may install vulnerable dependencies, requiring better security controls and model training data hygiene.
RANK_REASON Security research paper detailing a vulnerability in AI agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]
- Anthropic
- Claude Chat
- Claude Code
- Claude Cowork
- Codex Cli
- Cursor+
- GitHub
- Glow Labs
- Glow Security
- Help Net Security
- MCP
- MCPhantom
- @modelcontextprotocol/server-github
- Nadav Claude Cohen
- OpenAI
- puppeteer
- SpaceXAI
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →