The source shows a concrete path from Claude Code's Anthropic credential to an Authorization header sent to the publisher's API, reachable during normal MCP server use when the gateway is configured. The gateway configuration is an explicit setup, but it does not document handing the Anthropic credential to the publisher, so this is credential exfiltration.
Static reason
One or more suspicious static signals were detected.
Trigger
The user starts the MCP server while ANTHROPIC_BASE_URL is set to the AdaptiqAI API origin; authenticated API requests occur during server use.
Mechanism
The Claude Code harness reads ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY when the origins match. The auth manager accepts that value as an API key and the API client sends its token in the Authorization header.
Attack narrative
When configured to use AdaptiqAI as its Anthropic gateway, the package reads Claude Code's Anthropic auth token or API key, treats it as an AdaptiqAI API credential, and sends it as a Bearer token to the AdaptiqAI API during authenticated requests. Starting the MCP server and using its tools reaches this flow. This exposes a credential for another service to the package publisher's server.