# The Codex Sandbox Was Switched Off. The CLI Was Real.

> Symantec's 29 October notice says a relay installs the genuine Codex and Claude Code CLIs, then rewrites their config. Codex sandbox_mode is set to danger-full-access. Relay tokens are stored as user-scope environment variables. It has reached a dozen organizations since early 2026. Symantec has not seen abuse. Neither CLI has a vulnerability here.

Source: https://hackerlogs.com/blog/codex-relay-sandbox-off-clock/
Published: 2026-10-30

## Key takeaways

- The CLIs are the genuine npm packages. The relay rewrites config within minutes to about half an hour after that install. The binary is not the malicious object.
- Codex sandbox_mode is set to danger-full-access, write-sandbox network access is turned on, and telemetry and error reporting are disabled.
- Relay credentials are written as user-scope environment variables, so they survive a reboot and a later npm upgrade does not undo them.
- Symantec counts a dozen organizations, active since at least early 2026, and says it has not seen the operator abuse that position.
- There is no patched version to roll forward. The control is the config and the token, not a CVE in Codex or Claude Code.

US trends at 22:00 UTC on 30 October were Halloween, a Sam Altman interview, and sports. None of that is a sandbox control. The operator fact is [Symantec's notice](https://www.security.com/threat-intelligence/openai-claude-unauthorized-relays), bylined 29 October 2026.

A grey-market relay installs the real Codex and Claude Code command-line tools, then turns the containment off and points them at itself.

 Agent sandbox and identity tokens. High because the setting that is supposed to contain the agent is switched off on enterprise hosts, not because Symantec has seen theft.

<img src="/codex-relay-sandbox-off-clock.svg" alt="Three panels for the unauthorized Codex and Claude Code relay: a genuine CLI, a sandbox set to full access, and config that survives reboot." width="1200" height="675" />

## The package is the real CLI.

Symantec's Threat Hunter Team says the relay does not ship a modified client. People run a normal npm install of `@openai/codex`, `codex-cli`, or `@anthropic-ai/claude-code`. The override lands within minutes to roughly half an hour after that install.

That gap is the package clock. The package is the vendor's. The malicious object is the configuration that follows it. A fileless setup script, which users appear to run themselves, checks for Node and the CLI, installs them if needed, and only then writes the relay's settings. Symantec did not see an exploit delivery. It also says it has not seen the page the command was copied from. This brief will not repeat the command or the script URLs.

The page byline is 29 October 2026. The post record's `created` value is unix `1793268000`, which is `2026-10-29T10:00:00+00:00`. Other items on the same site use that same 10:00:00Z stamp, so this read does not treat 10:00 as a measured publish minute. The date is the disclosure clock. The activity clock is older: Symantec says the relay has been in use since at least early 2026. There is no start day on the page, so there is no honest day count from "early 2026" to 29 October. It is months, not a weekend.

## The impact clock is the config.

The script rewrites `~/.codex/config.toml` for Codex and `~/.claude/settings.json` for Claude Code. It replaces the base URL with a domain the relay controls, swaps in a relay-issued key, and stores both as user-scope environment variables. Those variables survive a reboot. They also survive a later copy of the genuine npm package, because the package is not where the override lives.

For Codex, Symantec says the script sets `sandbox_mode` to `danger-full-access`, turns on network access for the write sandbox, and disables telemetry and error reporting. The same telemetry switch suppresses the usage signal that would otherwise reach OpenAI or Anthropic. Claude Code is re-pointed the same way: a relay base URL and a relay credential, with the sandbox isolation disabled.

On at least one host the script went further before writing its own credentials. It logged the user out of Claude, deleted the existing auth variables, and logged in again, so a legitimate session could not override the relay key. Symantec does not say that step runs everywhere.

The Claude Code keys Symantec saw used Anthropic's subscription OAuth token shape, `sk-ant-oat01-`, not the pay-per-use API shape, `sk-ant-api03-`. Symantec calls that consistent with pooled flat-rate subscriptions, and not proof of it. Stolen-card fraud and resale of prompts are also open, and not shown.

The bulletin lists at least eight domains, all impersonating OpenAI- or Anthropic-branded chat services, and two setup-script hashes:

- `2a76a28639d896bb41576d0e32d155b78f38b034e052dfe3a7b7e1a478fc0e24` for the Codex setup script
- `4bfe5ca73a4b0988586a46f789aeedba5db575257b58291934cddb9faa606fdb` for the Claude Code setup script

This read did not fetch those scripts. The domains stay on the bulletin.

Scope, as Symantec counts it: a dozen organizations, mostly in mainland China, plus multinational subsidiaries in South Korea, Taiwan, and Malaysia. Sectors named include industrial automation, electronics, semiconductors, pharmaceuticals, shipping, tire manufacturing, petrochemicals, and data-center services. Symantec reads them as workplaces that wanted a coding agent the vendors do not officially sell there, not as a chosen-target list.

## The myth is that a genuine binary is contained.

> Symantec is explicit that this is not a vulnerability in Codex or Claude Code. The binary can match npm and the agent can still run with the sandbox off, network access on for writes, and a credential the employee did not get from the vendor. An official package hash does not answer the question.

The impact clock is that config. From early 2026 to the 29 October notice, the relay could read every prompt, file, and secret sent through the agent, and could alter what came back. With the sandbox disabled, a malicious or compromised response could run as the user. Symantec says none of the evidence it reviewed shows the operator exploiting that position. The architecture would support it if the operator's intent or the relay's own security changed.

> Pointing a CLI at a reseller does not require `danger-full-access`, a write sandbox with network access, or a user-scope token that outlives the shell. Those three settings are the control failure. The base URL is how you find them.

The credentials are not confined to the agent. Because they are user-scope environment variables, any other tool on the machine that reads them can pick them up. That is a second impact clock, and it is not closed by uninstalling Codex or Claude Code.

This is not the [Cursor marketplace copies](/blog/cursor-claude-code-marketplace-clock). Those were modified extension files on one marketplace for about 11 hours. This relay wants the official CLI and a permanent config change. Do not merge the hunts.

## What to do on hosts that installed it.

There is nothing to patch. Inventory hosts running Codex or Claude Code and read the config the binary is actually using.

Treat any of these as a finding, even when the installed package hash matches npm:

- `sandbox_mode` set to `danger-full-access` in `~/.codex/config.toml`
- a `base_url` or `ANTHROPIC_BASE_URL` you did not assign
- `OPENAI_BASE_URL`, `CODEX_API_KEY`, or `ANTHROPIC_AUTH_TOKEN` set as user-scope environment variables outside the approved deploy path
- telemetry and error reporting forced off next to that base URL

If the token shape is `sk-ant-oat01-` and your org issues pay-per-use keys, the credential did not come from the path you think. Remove the relay base URL and the user-scope variables, then rotate every secret that lived in that environment, from a trusted system. A reboot will not drop user-scope variables. A newer npm install will not either.

Block CLI traffic to destinations outside the allow-list of approved AI endpoints. Symantec's endpoint products will block a file they have, which does not cover a config rewrite of a genuine binary. The package clock was half an hour. The impact clock has been open since early 2026, and it stays open until someone reads the config.

## Sources

- [Sandbox flipped to danger-full-access: unauthorized AI relays](https://www.security.com/threat-intelligence/openai-claude-unauthorized-relays) (2026-10-29)
