MCP Protocol Flaw Lets Malicious Prompts Spread Between AI Agents

A newly detailed vulnerability class is exposing weaknesses in AI agents used by Google, JP Morgan Chase, Weaviate, Rapid7, and government agencies in France and the US. Independent researcher Syed Anas Mohiuddin demonstrates proof-of-concept attacks that exploit trust gaps in the Model Context Protocol (MCP), a standard for communication between AI apps and agents inside internal networks. The attacks use a special form of prompt injection that targets a specific agent, such as one for translation or data analysis, rather than the underlying LLM itself.

The problem stems from the architecture of these systems. Special-purpose agents often lack the guardrails needed to filter malicious instructions, and MCP servers store credentials for each agent. Because agents are built to trust other internal agents, a prompt that would be rejected by the LLM succeeds when passed along the chain. In many cases, well-crafted prompts aimed at the right agent result in server-side request forgery, causing web servers to make unauthorized network requests.

The consequences can be severe, including exfiltration of database contents and sensitive business or personal information. “AI agents give attackers a fresh set of connections to walk across,” says Douglas McKee, director of vulnerability intelligence at Rapid7. “Someone plants text in content, an agent will read it then pass it along to another agent as a normal delegated task.” Google and four other organizations have acknowledged such vulnerabilities over the past five months, and mitigation remains difficult given the structural nature of the flaw.

Read More at the original source →