Executive brief
Envoy AI Gateway, a tool used to manage and route traffic to AI services, is vulnerable to a message smuggling flaw. An attacker can send specially crafted requests that bypass security filters by using inconsistent capitalization in the data. This allows unauthorized commands to be executed on backend AI systems, potentially leading to unauthorized data access or misuse of AI tools.
Technical details
The vulnerability arises from a parser differential between the Model Context Protocol (MCP) specification and the Go-based JSON libraries used in Envoy AI Gateway (specifically `modelcontextprotocol/go-sdk` and `bytedance/sonic`). While JSON-RPC 2.0 requires case-sensitive matching, the gateway's internal unmarshaling is case-insensitive. An attacker can provide duplicate fields with different casing (e.g., 'name' and 'Name'). The gateway validates the compliant field but may overwrite it with the malicious value from the non-compliant field during internal processing. When the gateway re-serializes the message for the upstream backend, it produces a valid, canonicalized JSON-RPC request containing the smuggled malicious value, effectively bypassing any intermediate authorization layers. This affects 'tools/call', 'prompts/get', and 'resources/read' message types.
Affected products
- envoyproxy ai-gateway < 0.6.0
Timeline
- 2026-05-13: advisory: Initial GHSA disclosure
- 2026-05-19: patched: Version 0.6.0 released