Skip to content
Adversa AI·Indexed by this site · Originally published· Event →

November 2026 MCP Security Resource Roundup

Top MCP security resources — November 2026

AI summary

Adversa AI's monthly compilation includes 9 MCP security resources, of which 5 are MCP vulnerabilities, covering forgeable session IDs in Envoy AI Gateway, server-everything leaking environment variables due to wildcard CORS and no authentication, Basic Memory subjected to DNS rebinding, FastMCP OAuth clients lacking issuer binding, and the MCP TypeScript server SDK mixing messages across shared instances, with the respective fixed versions given for each. The Log-To-Leak prompt injection technique achieved attack success rates of 62.64% to 100% across 18 model-and-server combinations, with target task completion changing by no more than 6.63 percentage points in 16 combinations. In an MCP test environment with 24 adaptive payloads, keyword and regex defenses each let through 62.5%, while the HMAC provenance-based scheme blocked all payloads within a controlled execution environment.

Related papers
Read originaladversa.ai