The CLI was the real one
Symantec found a grey-market relay at about a dozen organizations that installs the genuine Claude Code and Codex CLIs, then rewrites their configuration with one pasted command. Every prompt, file and secret now goes to the relay's servers, the agent's sandbox is set to full access, and telemetry is off. Google's LLMjacking report this week puts relays like it in a market for stolen and resold model access. In Bromure Agentic Coding, the agent's model connection is set on the Mac, the VM only holds stand-in keys, an allow-list firewall refuses the relay's servers, and a rewind removes the edit.
The agent in this attack is the genuine one, installed from npm. One pasted command edits two settings files, and from then on every prompt, file and secret the agent handles goes through someone else's servers.
On October 29, Symantec's Threat Hunter Team published Sandbox Flipped to "Danger-Full-Access", a report on an unauthorized relay service that resells access to OpenAI Codex and Anthropic Claude Code. On November 3, Google's threat intelligence group and Mandiant published research on LLMjacking, the theft and resale of access to AI models. The Hacker News and BleepingComputer covered the two together. In Google's cases, stolen access ran up more than $100,000 in charges in a single day, and up to $1 million at some organizations before anyone revoked it. Google describes "LLM mixers" that pool and resell that access, one of them offering an unlimited plan for about $60 a month. Symantec's relay is the developer-facing end of the same market.
Symantec found it at about a dozen organizations, mostly in mainland China, plus subsidiaries of companies based in South Korea, Taiwan and Malaysia. The sectors were industrial automation, semiconductors, pharmaceuticals, shipping, petrochemicals and data center services. Symantec reads them as workplaces where employees went looking for tools the vendors don't sell in their market, rather than chosen targets. It has been active since at least early 2026.
Two commands, half an hour apart
Every case Symantec saw started the same way:
- The employee installs the real CLI with
npm install -g. Nothing about that install is malicious, which is why it looks routine to the user and to most monitoring. - Between a few minutes and about half an hour later, they paste a one-line
command that downloads a bash or PowerShell script from
vibe.aiok[.]meand runs it in memory. The script prints its progress in Chinese. Symantec thinks users copy the command from the relay or a reseller.
The script edits ~/.codex/config.toml for Codex and
~/.claude/settings.json for Claude Code:
- Base URL. The address the agent sends its model requests to now points
at a relay domain. Symantec lists at least eight, including
claudechn[.]com,sharedchat[.]cc,rawchat[.]cnandnuoda[.]vip. - Key. The user's key is replaced with one the relay issued. On one host
the script also ran
claude /logoutand deletedANTHROPIC_AUTH_TOKENandANTHROPIC_API_KEY. The relay's Claude keys carry thesk-ant-oat01-prefix of subscription sign-in tokens. - Persistence. The new URL and key are saved as user-scope environment variables, so they survive a reboot, and any other program on the machine can read them.
- Sandbox. Codex's sandbox mode is set to
danger-full-access, and the write sandbox gets network access. - Telemetry. Telemetry and error reporting are switched off, which, Symantec notes, "would also suppress any usage signal."
Symantec is clear that this is "not a vulnerability in Codex or Claude Code." The tools do what their configuration tells them to.
What the relay holds
Symantec found no sign that the operator has misused its position, but the report spells out what the position allows. A relay there "sees every prompt, file, and secret an employee sends to the agent" and "can alter or fabricate every response it returns."
A coding agent's requests carry most of a codebase over a day of work:
source files it reads, diffs it writes, logs it pastes, the .env it opens
while debugging. With the sandbox at full access and network access on, a
response that tells the agent to run a command runs it with the user's full
privileges. The relay can make the agent run a command by writing its next
instruction.
Two settings files and two environment variables separate "the agent talks
to Anthropic" from "the agent talks to a stranger." The grey-market service
asked users to make that edit themselves.
A malicious postinstall script, a README command an agent pastes for you,
or a prompt injection in a page the agent reads can make the same edit
without asking. Google's report says the same: attackers clone repositories
to hunt for keys and developer credentials, then resell the access.
The same command in Bromure Agentic Coding
Bromure Agentic Coding runs Claude Code, Codex, Grok, Kimi and omp in Ubuntu VMs on your Mac. Each machine is a VM of its own, and every connection it makes goes through a proxy on the Mac, which holds the real credentials and records what happened.
The model connection is set on the Mac
In Bromure, the agent's model access doesn't come from a config file the agent can edit. It comes from Preferences › Models on the Mac, the Models board, where you register each source once: an Anthropic, OpenAI, xAI or Moonshot subscription or API key, Amazon Bedrock, OpenRouter, a custom inference server, or an on-device model on the Mac's GPU. Each agent gets a model from a menu, and each machine can inherit the board or override it.
That covers most of what sends people to a relay. A cheaper model, a different provider, a model that runs offline: each is a menu entry, set up on the Mac and signed in through the provider's own page in your Mac's browser. You get a working agent without pasting a stranger's install command.
The VM holds stand-ins, not keys
Inside the VM, ANTHROPIC_API_KEY holds a fake that starts with
sk-ant-api03-brm-. A Claude or ChatGPT subscription works the same way: the
real sign-in tokens stay on the Mac and the VM runs with a placeholder. The
proxy swaps the fake for the real key only on requests to that provider's own
hosts.
Run the relay's script in a Bromure machine and look at what it gets:
- The
claude /logoutand the deleted variables remove a placeholder. Your real sign-in is on the Mac, out of the VM's reach, and every other machine keeps using it. - The user-scope environment variables that other programs can read hold the relay's own key. None of your credentials are in them.
- A config that changes the base URL but keeps your key, which is the variant a thief would use, sends Bromure's fake to a host it wasn't minted for. The compromise detector catches that. The request is blocked with HTTP 451 before a byte leaves the Mac, the VM pauses, and an alert titled This environment may have been compromised names the credential, the host it belongs to and the host it was sent to. Shut down is the default answer.
The firewall knows where models live
Symantec's first detection tip is to flag "traffic from AI CLI tools to destinations outside an allow-list of approved AI service endpoints." In Bromure that is a setting. Each machine's Guardrails pane has an Outbound connections firewall: rules by host, IP range, protocol and port, checked top to bottom, with Unmatched traffic set to Deny for an allow-list:
allow web anthropic.com
allow web openai.com
allow web github.com
allow web registry.npmjs.org
default deny
A host rule covers the domain and its subdomains. The Mac enforces the rules twice: the VM's virtual network switch checks every connection by destination IP and by the hostname seen in DNS, for every protocol, and the proxy checks again by TLS server name. The VM has no other route out.
On that machine, the rewritten settings.json points Claude Code at
claudechn[.]com and the connection is refused on the Mac. The agent's next
turn fails with a connection error in front of you, the Firewall engine
logs a red "deny" row for the relay host in the Security Timeline, and
on enrolled Macs the organization gets an egress.firewall event with the
host, port and action. No prompt, file or secret reaches the relay.
The boundary isn't in the config
The relay set Codex's sandbox to danger-full-access because on a laptop
that setting is the main thing between the agent and the rest of the
machine. In Bromure, the VM draws that line. At full access, the agent reaches everything in an Ubuntu guest
that holds stand-ins, and its network still runs through the Mac's firewall
and proxy. No file inside the VM changes that.
Bromure keeps its own record on the Mac, and the agent's telemetry switch can't reach it. Every firewall verdict, credential swap and blocked request is in the Security Timeline, which you can filter and export. The Protections table in the same window marks which machines have firewall rules, so you can spot a machine left open.
The edit comes out with a rewind
The script persists through files in the home folder: two settings files and the environment variables it saved for the user. Bromure takes a copy-on-write checkpoint of each machine's home on every boot. Shut the machine down, choose Rewind home…, pick the boot before the paste, and the next start comes up with the agent's original settings. The home as it was is kept as a new point, so you can still inspect what the script changed.
Models set on the Mac
Providers, subscriptions, Bedrock, OpenRouter, custom servers and on-device models live on the Models board. Switching is a menu choice.
Stand-ins in the VM
The agent's key is a fake. A fake sent to a relay gets a 451 and pauses the VM. A logout in the VM leaves your real sign-in on the Mac.
An allow-list for models
With unmatched traffic set to Deny, the relay's domains are refused on the Mac, at the network switch and at the proxy.
On record, then rewound
The Security Timeline logs the blocked host. Rewind home restores the agent's settings from before the paste.
Check where your agents send their requests
Symantec's checklist is a good place to start on any fleet: look for
base_url or ANTHROPIC_BASE_URL changes in the CLIs' configuration,
sandbox_mode = "danger-full-access" in ~/.codex/config.toml, user-scope
OPENAI_BASE_URL, CODEX_API_KEY or ANTHROPIC_AUTH_TOKEN variables set
outside your deployment process, and a curl | bash or iwr | iex shortly
after an npm install of an AI CLI.
A checklist finds the edit after it happened. In Bromure Agentic Coding, the place an agent sends your code is decided on the Mac, the firewall knows which hosts are allowed, and a rewritten settings file inside the VM leaves the agent with a refused connection and you with a red row in the timeline.
For an agent whose configuration was rewritten by a web page it read, see The web page rewrote the agent's config. For what happens to AI keys copied onto other platforms, read The key had too many homes.