Urgent.News

What's breaking now, across hundreds of outlets.

Tech

Securing Model Context Protocol (MCP) Servers: Stopping Prompt Injection and Secret Leaks

Securing Model Context Protocol (MCP) Servers: Stopping Prompt Injection and Secret Leaks The Model Context Protocol (MCP) has rapidly become the standard interface connecting AI agents and IDEs to local developer tools, databases, and APIs. Instead of building custom tool interfaces for every LLM, developers can expose a standardized JSON-RPC server over standard input/output ( stdio ) or…

Model Context Protocol (MCP) servers have become widely adopted for connecting AI agents to local developer tools, databases, and APIs. However, this standard interface presents new security challenges. Three primary vulnerability areas must be addressed to secure MCP implementations:

1. Tool docstring and parameter schema injection: When an AI agent connects to an MCP server, tool definitions including names, descriptions, and parameter schemas are returned in JSON format. Attackers can manipulate these descriptions to trick the model into executing unauthorized actions. The solution is to implement schema sanitization gates that validate parameter schemas strictly and remove any malicious content from tool descriptions before they are presented to the AI model.

2. Unrestricted stdio child process spawning: MCP servers typically communicate using standard input/output (stdio) when launched from an IDE or agent launcher. If the MCP server runs with the same privileges as the parent process, it can gain access to all local files, environment variables, and sensitive credentials. To mitigate this, adopt the principle of least privilege by restricting filesystem access to specific directories, sanitizing environment variables, and mounting read-only paths for tools that only require data inspection.

3. Tool argument injection and path traversal: AI models generate arguments dynamically when invoking MCP tools. Adversaries can exploit prompt injection vulnerabilities in upstream data sources to inject malicious inputs through these arguments, potentially leading to file system manipulation or database queries. To prevent this, enforce deterministic input validation by verifying canonical paths and rejecting any raw SQL or shell command generation by the model. Instead, utilize parameterized queries and strictly typed domain actions.

Rewritten by urgent.news; may contain errors — check the original.

Source: dev.to

Read the original at dev.to →

Report a problem with this story

More in Tech

What is an agent execution receipt standard?

The short answer An agent execution receipt standard is a published specification that defines, for a record of one AI agent tool call, four things exactly: what the record contains, how the…

  • Agent execution receipt standard defines essential details of AI agent tool calls
  • AER-1 is a published IETF Internet-Draft exemplifying the standard
  • Receipt record includes execution details, commitment, verifier outcomes, and conformance corpus

Found a Cracked Copy of Your App? A Practical Response Plan for Indie Devs

Originally published on the PermitCore blog . You found your paid desktop app on a warez site. Someone posted a keygen. A GitHub repository contains a modified build with your payment screen removed.

  • Determine type of infringement: cracked binary, keygen, re-upload, clone, or impersonation.
  • Escalate issue to lawyer if platform rejects notice or fails to respond.

Human, for Now: The Adjective Is Changing Sides

Two years ago that trailer was the interesting part. It flagged the unusual thing. I think we're close to the point where the interesting line is the other one: what did the human do?

  • "Human" named Word of the Year by Cambridge Dictionary for AI's impact on perception of humanity
  • Proposed metadata system to track human-AI collaboration levels in content creation

More from Sunday 22 November →