CodexAPI.pro Confirmed for Spoofing GLM as Claude and Security Risks

CodexAPI.pro advertises that it provides Anthropic's Claude models, but network packet analysis has confirmed that it actually routes requests to Zhipu AI's GLM models and disguises the models by manipulating SSE metadata (model spoofing).

Model Spoofing and Protocol Translation

• The proxy converts Anthropic API requests into OpenAI-compatible format and forwards them to Chinese Zhipu AI endpoints.
• OpenAI tool call IDs in the chatcmpl-tool-* format are exposed in the response stream, revealing that the actual backend is GLM.
• False model information such as Openai-Model: claude-opus-5 is inserted into HTTP headers to bypass CLI mismatch errors.

Security Risks and Installation Scripts

• The installation script sets approval_policy="never" and wildcard permissions, removing the Human-in-the-loop approval process.
• It modifies system configuration files such as ~/.zshrc, registers background daemons, and obtains clipboard access permissions.
• It advertises unreleased models (such as GPT-5.6), while the actual cost reduction structure involves passing off low-cost models (GLM-4) as frontier models to generate profit margins.

Data Governance and Response

• User code is transmitted to Chinese servers via third-party infrastructure rather than Anthropic/OpenAI, posing data leakage risks.
• Numerous gray market services disguise low-cost open-weight models as frontier models; users should test them only in isolated environments such as non-privileged Docker containers.