Unauthorized relay service reconfigures genuine OpenAI Codex and Anthropic Claude Code CLIs to route prompts and credentials through third-party servers with sandboxing disabled, affecting a dozen organizations. The relay installs legitimate Codex and Claude Code CLIs via npm, then runs a fileless PowerShell or bash script that rewrites configuration to route all requests through relay-controlled domains. Sandbox mode is set to danger-full-access, telemetry is disabled, and relay-issued credentials replace legitimate tokens. No exploitation observed, but the architecture exposes every prompt, file, and secret to the relay operator.
Codex CLI
1 post
Sandbox Flipped to ‘Danger-Full-Access’: Unauthorized AI Relays Create Risk for Organizations