Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsIncident

Symantec: an unofficial relay for Codex and Claude Code rewrote their settings, turned the sandbox off and routed every prompt through its own servers

Symantec's Threat Hunter Team reported on October 29 that an unauthorized relay service, seen since at least early 2026, gives users access to OpenAI's Codex and Anthropic's Claude Code command-line agents, mainly in mainland China, where the vendors do not officially sell. About a dozen organizations were affected, mostly in mainland China, plus subsidiaries of multinationals in South Korea, Taiwan and Malaysia, in industrial automation, electronics, semiconductors, pharmaceuticals, shipping, petrochemicals and data centers. The user installs the genuine tool from npm, then runs a one-line command copied from a lookalike domain. The script points the tool at the relay's servers, sets Codex's sandbox to danger-full-access, turns off Claude Code's sandbox isolation, swaps the user's credentials for relay-issued keys, disables telemetry and error reporting, and stores the new values as user environment variables that survive a reboot. Symantec says this is not a vulnerability in either tool, and that it saw no evidence the operator abused the access, but the relay could see every prompt, file and secret sent through the agent and could alter its answers, which with the sandbox off would run with the user's full rights. The infrastructure rotated across at least eight domains made to look like OpenAI or Anthropic chat services.

Why it matters

A coding agent is only as trustworthy as the server its settings point at. Watch for a changed base URL, a full-access sandbox and AI keys set outside your approved deployment, and treat any "one line to install" command for an AI tool as code you are running blind.

Sources

Read it with the rest of the news in the Learning Center