The attack surface for AI coding agents has expanded beyond the codebase itself. Developers using Claude Code, Cursor, Windsurf, and Cline frequently clone repositories containing SKILL.md, .cursorrules, CLAUDE.md, and mcp.json files. While these are often treated as passive documentation, they function as executable control planes. When an agent has bash and filesystem access, a malicious skill file can instruct it to read .env.local, exfiltrate credentials, or execute arbitrary shell commands without explicit user confirmation.

Twelve Threat Vectors for Agent Configs

Aviv Gerzhue has released VibeShield v2.1, a security scanner that introduces 12 static rules dedicated to auditing AI agent configurations. Designated as VSB-MCP-001 through VSB-MCP-012, these rules operate alongside 80 existing Next.js, Supabase RLS, and TypeScript application security checks. The tool specifically targets the unique vulnerabilities introduced by agentic workflows, where instruction files can override safety boundaries.

Obfuscation and Exfiltration Tactics

The new rules detect sophisticated obfuscation techniques, including zero-width and bidirectional Unicode characters (U+200B, U+202E) hidden within SKILL.md or .cursorrules files. These invisible instructions do not appear in standard code reviews but can hijack agent behavior. Additionally, the scanner flags prompt override directives that instruct the agent to ignore prior safety boundaries or suppress warnings, effectively disabling guardrails set by the developer.

Credential Theft and Remote Execution

VibeShield v2.1 identifies credential and .env exfiltration directives by catching skill instructions that reference sensitive paths like ~/.ssh, ~/.aws, or token files when combined with outbound network tools such as curl, wget, or fetch. The tool also flags remote shell pipe execution patterns (curl | sh) in agent workflows or MCP hooks, which allow arbitrary remote scripts to run directly in the local environment. Furthermore, it audits mcp.json and .cursor/mcp.json for overly permissive flags like --allow-all or --dangerously-skip-permissions, as well as unpinned npx servers that could introduce supply chain risks.

Rapid Pre-Flight Scanning

To mitigate these risks, developers can scan any repository in seconds before opening it in their IDE. Running npx -y @vibefoundry/vibe-shield@latest scan . checks for immediate threats, while npx -y @vibefoundry/vibe-shield@latest surface . maps the complete attack surface, including public API routes, webhooks, database tables, and agent skill files. The full suite, which also includes GhostQA for browser testing and Unslop for UI auditing, is open-source under the MIT license.

Key Takeaways

  • Agent instruction files (SKILL.md, .cursorrules) are executable control planes, not passive documentation.
  • VibeShield v2.1 introduces 12 specific rules (VSB-MCP-001 to VSB-MCP-012) to detect prompt injection and exfiltration.
  • Hidden Unicode characters and prompt override directives can bypass standard code review processes.
  • Pre-flight scanning with npx allows developers to audit repositories before granting agent access to local files.

The Bottom Line

If you are treating agent configuration files like READMEs, you are leaving your local environment exposed to silent exfiltration. Static analysis of control planes is now a mandatory step in any secure agentic workflow.