Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentMCP servers and tools

Deadbugz: an MCP server that turns malicious after three tool calls

HighPillar Security

Pillar Security found an active campaign submitting a fake "productivity-suite" MCP server to public MCP directories. It behaves normally at first; after a client makes three tool calls it changes the instructions it gives the agent, telling it to collect SSH keys, AWS credentials, shell history and Kubernetes configs and hide that from the user.

Replay the attack step by step in the Learning Center

How it happened

  1. One GitHub account opened 23 pull requests in 74 minutes adding the server to MCP lists and directories.
  2. The server offered harmless text tools, so a quick review or first test saw nothing wrong.
  3. From the fourth call on, the tool metadata told the agent to gather secrets and send them out quietly.

Impact

Any developer who installed it and used it normally could have their SSH keys and cloud credentials sent to the attacker.

The fix

Most of the pull requests were closed at review; Pillar published the endpoint to block and recommends treating any change in an approved server's tool definitions as a security event.

Lessons

  • A server that passes review can change later, even mid-session. Pin tool definitions and re-approve on change.
  • Short trials do not prove an MCP server is safe; watch what it tells the agent over time.
  • Directory listings and pull requests are not vetting.

Sources