Top 25 MCP Vulnerabilities Reveal How AI Agents Can Be Exploited

SecurityWeek reports a new ranking of Model Context Protocol (MCP) weaknesses, highlighting risks from prompt injection to command injection and outlining a practical security framework for agentic AI. Developed by Anthropic and released as an open standard in 2024, MCP is described as the de facto (not absolute) standard for a safe interface between AI agents and data sources.

The article notes that prompt injection remains the top vulnerability, combining critical impact with trivial exploitability, while the MCP Preference Manipulation Attack (MPMA) is ranked at #24 with low impact and very complex exploitability. AI-specialist firm Adversa has published the analysis, described as the most comprehensive to date, and the piece indicates OWASP is planning its own Top Ten for MCP to be aligned with broader security frameworks, though not yet available.

The ranking is produced with a 40% weight on impact, 30% on exploitability, 20% on prevalence and 10% on remediation complexity, and the piece also includes a mitigation checklist and a three‑month timeline beginning with authentication on exposed endpoints. According to SecurityWeek, Polyakov of Adversa says the document will be updated monthly or as new incidents or CVEs require. 23 September 2025