Hundreds of MCP Servers Expose AI Models to Abuse, RCE

Hundreds of publicly exposed Model Context Protocol (MCP) servers are misconfigured, risking sensitive data leakage, context-poisoning of LLMs, and remote code execution; researchers estimate ~15,000 MCPs exist, ~7,000 are web-accessible, a couple hundred are wide open to local-network connections, and about 70 have serious vulnerabilities. The report describes neighborjacking and command-injection risks, notes no evidence of malicious MCPs found yet, and recommends immediate mitigations (inventory/scanning, input validation and sanitization, restricted filesystem access, strong access controls, and using secure transports).