The Agent in the Middle: How AI-Mediated Access Reshapes Web Vulnerability Exploitation

Summary

Researchers built VELA, an executable benchmark of 44 vulnerabilities across twelve Web application families, and tested three commercial AI agents against agent-facing interfaces derived from legitimate workflows. The agents realized every attack that direct interface invocation realized in the reference configuration, with 91% of vulnerabilities remaining exploitable, and interface design affected exploitability more than agent choice, moving retention from 39% to 95%.

Related items

• HighCVE-2026-118378: Tabularis SQLite driver SQL injection through MCP describe_table toolSimilar attack · NVD/CVE Database
• HighGHSA-m65p-9j7q-6w33: MLflow vulnerable to arbitrary file read on unauthenticated servers via prompt model version sourceSimilar attack · GitHub Advisory Database
• CriticalGHSA-26p8-2jq9-3vq9: MLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifact modulesSimilar attack · GitHub Advisory Database
• MediumGHSA-w4fh-qvv9-3v23: Model Context Protocol Python SDK: Server bearer-token authentication accepts tokens issued for a different resource serverSimilar attack · GitHub Advisory Database
• MediumGHSA-5h93-6whr-6q8j: Model Context Protocol Python SDK: HTTP client transports followed cross-origin redirects carrying custom headers and bodiesSimilar attack · GitHub Advisory Database