July 27: AI agents and coding assistants are becoming the fastest-growing part of the enterprise toolkit and, according to new Checkmarx research, one of the least understood risks. Researchers found that a single, booby-trapped text file is all it takes to hijack tools built on the Model Context Protocol, the standard powering many of today’s AI integrations. No malware. No exploit. Just a config file sitting quietly in a repository, waiting to be trusted.
The researchers proved the risk is real by reproducing a remote code execution vulnerability in a popular AI-powered security scanner, meaning the very tool meant to catch threats became the mechanism for delivering one.
Commenting on the broader trend of AI agent risk outpacing organizational awareness, Dor Tumarkin, Director of Security Research at Checkmarx Zero, noted:
“As AI research continues to create new capabilities in autonomous agents, and industry leaders are under pressure to gain their benefits, it creates a perfect recipe for safety and security risks.” He added that organizations need to understand “the capabilities of their AI systems, the safety of their AI supply chain, and the limitations of the security and safety controls they rely on to contain AI agents.”
What is MCP? The Model Context Protocol is an open standard that lets AI agents, coding assistants, and IDEs connect to external tools, data sources, and servers — essentially giving AI systems a common way to “plug in” to the software developers already use. It’s been rapidly adopted across the industry for exactly that reason, but that same openness means any tool that automatically trusts and processes an MCP configuration file is also automatically exposed if that file has been tampered with.