More than 7,000 of over 15,000 Model Context Protocol servers, which enable artificial intelligence app access to data outside their training models, around the world are exposed to the internet, hundreds of which are impacted by the "NeighborJack" vulnerability that allows access to anyone within the same local network, Infosecurity Magazine reports.
Nearly 70 of the MCP servers also had unnecessary permissions, uncontrolled input handling, and other significant vulnerabilities, a report from Backslash Security showed. Attackers could leverage such issues not only to facilitate total host machine hijacking but also context poisoning intrusions that involve large language model data tampering, according to researchers, who emphasized that none of the examined MCPs have been malicious. Organizations have been advised to mitigate potential exploitation of MCP servers by restricting local network interface access, limiting file system access, conducting external input validation, adopting robust authentication and access controls, and preventing internal log exposure in AI responses.
Nearly 70 of the MCP servers also had unnecessary permissions, uncontrolled input handling, and other significant vulnerabilities, a report from Backslash Security showed. Attackers could leverage such issues not only to facilitate total host machine hijacking but also context poisoning intrusions that involve large language model data tampering, according to researchers, who emphasized that none of the examined MCPs have been malicious. Organizations have been advised to mitigate potential exploitation of MCP servers by restricting local network interface access, limiting file system access, conducting external input validation, adopting robust authentication and access controls, and preventing internal log exposure in AI responses.
