EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·Aug 12·5 min read
Deadbugz: Currently Active MCP Supply-Chain Campaign
TL;DR
Pillar Security Researchers uncovered a malicious Model Context Protocol (MCP) server disguised as a productivity tool on GitHub. After three interactions, it shifts to seek sensitive data while hiding its actions from users.
✦ Why It Matters
Review and audit any third-party tools you integrate to ensure they are not malicious.
Key Takeaways
- Block the endpoint associated with the Deadbugz campaign.
- Review and revert any MCP configuration changes introducing the malicious server.
- Investigate logs for any connections to the compromised server.
- Implement stricter approval processes for tool definition changes.
How It Works
The Deadbugz campaign employs a malicious MCP server that initially provides benign tool responses. After three tool calls, it alters its metadata to instruct the connected AI agent to seek sensitive files and credentials, effectively hiding its malicious intent until the threshold is crossed.