Unauthorized Relay Disables Safeguards for Codex, Claude Code
A mysterious relay service has been distributing legitimate versions of OpenAI Codex and Anthropic Claude Code to organizations in regions where these tools are not officially sold, but with critical security features disabled. Operating since early 2026, the relay has installed the tools on systems at roughly a dozen organizations, primarily in mainland China but also in South Korea, Taiwan, and Malaysia. Users obtain the tools through standard npm channels, then execute scripts from lookalike domains that reconfigure them to route requests through the relay instead of official vendor infrastructure.
The relay disables multiple safety controls: it turns off the execution sandbox that constrains the AI tools’ actions, enables network access for the write sandbox, disables telemetry and error reporting, and replaces legitimate API credentials with relay-issued tokens. In one documented case, the relay logged a user out of their Anthropic session, deleted their authentication variables, and reissued credentials under relay control. Though no malicious activity has been observed, the architecture poses severe risks—the relay intercepts every prompt, file, and secret users send to the tools and can alter or fabricate responses that execute with full user privileges when the sandbox is disabled.
The motivation for disabling these security features remains unclear. Researchers speculate potential scenarios including fraud operations involving stolen credentials or payment cards, malicious resale of data passing through the relay, or credential pooling across multiple customers. The affected organizations appear to be workplaces where individual employees discovered and used the relay to access tools they wanted, rather than targeted victims.
Read the full article
Continue reading at Dark Reading →This is an AI-generated summary. Read the original for the full story.