The backdoor outlived the package
On November 9, attackers used Ethersphere's own release automation to publish bee-js 13.3.0 to npm and a signed, notarized Swarm Desktop 0.56.2. The npm version was live for 75 minutes, with valid provenance. Its install script stole Swarm wallet keys and planted a backdoor in the home folder that removing the package does not remove. The advisory says any machine that installed it is fully compromised. In Bromure Agentic Coding, the agent's npm never sees a version that young, and if a machine does run it, Rewind home puts the home folder back to how it was before the install.
npm pulled the bad version 75 minutes after it went up. That didn't help anyone who had already installed it, because by then the payload had put its backdoor in their home folder. The advisory's fix is to rebuild the machine.
Early on November 9 (UTC), someone used Ethersphere's own release pipeline
to ship malware. Ethersphere maintains Swarm, a decentralized storage
network. Its JavaScript library, @ethersphere/bee-js, is what Swarm
Desktop, swarm-cli and the Swarm dashboard use to talk to a node.
Socket and
SafeDep
published the technical details, and The Hacker
News
covered it the same day.
Socket's timeline, in UTC:
- Nov 8, 23:30: the attackers point an Ethereum smart contract at their
command server,
193.57.9.155:4912. - Nov 8, 23:59: an unsigned commit, authored as the release bot
bee-worker, renames bee-js'spreparescript topostinstalland ships thescriptsfolder in the package. - Nov 9, 00:06: a GitHub release
v13.3.0triggers the Publish on npm workflow. Nothing in it requires an approval or checks that the commit sits on a protected branch. The job finishes in 73 seconds. - 00:09:50: npm serves
@ethersphere/[email protected]aslatest. - 00:41–00:44: a workflow the attackers started by hand builds Swarm Desktop 0.56.2, signs it with the project's Apple and Windows certificates, gets it notarized by Apple, and the app's updater starts offering it.
- 01:24:58: npm removes 13.3.0. GitHub advisory GHSA-hh7j-5m3j-23pf follows 29 seconds later.
The package came with valid provenance. Socket verified the attestation
offline: it names the real workflow, the real commit and the real publish
environment. In Socket's words, provenance "proves where a package was
built, not that the source was reviewed." Two GitHub accounts were
compromised, the release bot and a long-time maintainer who could still
force-push to an unprotected branch. Two days earlier, workflows in two
other Ethersphere repositories had tried to dump every CI secret to
webhook.site. How the attackers got into either account is still
unexplained.
What the install did
Version 13.3.0 was the first of bee-js's 178 versions to declare an install
hook. The hook calls https://api.frameworkmetrics.xyz/v1/analytics with an
X-Ethersphere: 1 header, and runs whatever JavaScript comes back with
Node's real require and process, as the user who ran the install.
Without the header, the server sends a decoy. The code comments say the
script "must NEVER fail an install."
You didn't have to ask for bee-js to get it. swarm-cli 3.6.0 and 3.7.0,
bee-dashboard 0.37.x, @snaha/swarm-id, bee-wasm-js, the
file-manager-lib package and any new create-swarm-app project all
accept version ranges that include 13.3.0. For those 75 minutes, a fresh
install of any of them resolved to the malicious version.
The second stage looks for Swarm Desktop's data folder and uploads two
files to /v1/swarm/report: config.yaml, which holds the node's wallet
password in plain text, and keys/swarm.key, the encrypted Ethereum
keystore it unlocks. With both, the attackers can move the node's tokens,
its chequebook and its stake.
Then it moves in. On Linux and macOS it writes a Go remote shell, which
SafeDep calls evjawbreaker, to a hidden ~/.solidity folder, and starts it
with a systemd user unit named solidity-langsupport-runner.service or a
LaunchAgent named com.solidity.langsupport.runner. The implant reads its
server address from the Ethereum contract, so there is no fixed domain to
block, and the attackers can move it with one transaction. On Windows the
payload installs a signed ScreenConnect client as a SYSTEM service, after a
UAC bypass when it can.
Removing the package removes nothing
The GitHub advisory is short: "Any computer that has this package installed or running should be considered fully compromised." The cleanup list that follows is long:
- Search lockfiles, caches, internal mirrors and CI logs for
@ethersphere/[email protected], including indirect copies. - Treat any machine or CI job that installed it with scripts enabled between 00:09:50 and 01:24:58 UTC as compromised, even if the install failed.
- From a clean computer, rotate every password, token and key the installing user could read, including cloud credentials reachable through CI OIDC trust.
- Move the Swarm node's funds, chequebook and stake to a new key.
- Delete
~/.solidity/and the systemd unit or LaunchAgent. Updating, reinstalling or removing the package or the app doesn't touch them. - Where possible, rebuild the machine from a trusted image.
On a developer laptop, the last two items take most of a day. "Every
password, token and key the user could read" covers ~/.npmrc,
~/.config/gh, ~/.aws, ~/.ssh, every .env in every checkout, the
browser's cookies and whatever the agent pasted into its shell history.
And rebuilding from a trusted image means reinstalling the laptop.
Coding agents make that day more likely. Scaffolding a Swarm app, fixing a
broken build or bumping a dependency all end in npm install, and agents
run it without stopping to read which versions it picked. An agent told to
"try the Swarm CLI" between 00:09 and 01:24 would have installed 13.3.0
through swarm-cli's version range.
The same 75 minutes in Bromure Agentic Coding
Bromure Agentic Coding runs Claude Code, Codex, Grok, Kimi and omp in Ubuntu VMs on your Mac, one VM per machine. Every package download goes through a proxy on the Mac, and the agent's credentials inside the VM are stand-ins.
npm never saw 13.3.0
The supply-chain age gate is on by default on every machine, with a
two-day minimum. The proxy removes versions younger than the cutoff from
npm's version listing and points latest at the newest version that's old
enough. When the agent asked for swarm-cli and npm resolved bee-js's
range, the newest bee-js it could see was 13.2.0. The install went through
and the build worked, with no error for the agent to chase.
If the agent had pinned @ethersphere/[email protected] by name, the download
would have failed with HTTP 451 and a body npm prints word for word:
Bromure Supply-Chain Security blocked this request: npm package
@ethersphere/[email protected] published 40 minutes ago — policy requires
2 days minimum
npm removed the version after 75 minutes. A two-day gate leaves 46 more hours for the reports to come in. We covered the same pattern on PyPI in The release was two hours old.
The install hook had nobody to run it
npm inside a Bromure machine runs install scripts only for packages on its
allow list, and Bromure's managed ~/.npmrc puts only the agent CLIs it
ships on that list. bee-js's new postinstall would have been skipped.
The machine's Supply Chain pane adds a second, host-side layer:
Strip preinstall / install / postinstall / prepare from npm tarballs on
the fly. The proxy removes those hooks from each tarball before it
reaches the VM, so even an npm someone reconfigured inside the machine
has nothing to run. With a socket.dev key, Block compromised packages
refuses packages socket.dev flags as malware, at the download.
Say it ran anyway
Teams that work on Swarm itself would add their own packages under Exempt packages, so they can test a release the minute it ships. Take that worst case: the agent installs 13.3.0 with scripts on, and the payload runs in the VM.
The wallet isn't there. The second stage looks for Swarm Desktop's data
folder. On a Mac, that's ~/Library/Application Support/Swarm Desktop.
The VM has its own Linux home and sees only the folders you share with it.
Swarm Desktop's data folder is not one of them.
The tokens are stand-ins. The advisory says to rotate everything the
installing user could read. In a Bromure machine, that user could read
ghp_ values GitHub never issued, an NPM_TOKEN that starts with brm_,
an AWS secret key that's fake, and an SSH_AUTH_SOCK that asks the Mac for
signatures and never holds a private key. The proxy on the Mac swaps each
stand-in for the real value only on requests to the host it belongs to.
Copied to the attacker's laptop, they fail authentication.
A stand-in sent to the wrong host trips the alarm. If the backdoor packs the machine's tokens into a request to any host they weren't minted for, the proxy blocks it with HTTP 451, pauses the VM, and opens an alert titled This environment may have been compromised. It names the credential, the host it belongs to and the host it was heading for. Save for Investigation exports the VM's disk and home for your incident response before the machine shuts down.
The backdoor needs a way out. Every connection the VM makes passes
the Mac's virtual switch and proxy, and the machine's Guardrails pane
holds an egress firewall that takes rules by host, IP range, protocol and
port. With Unmatched traffic set to Deny, a machine that does web
development reaches the registries and APIs on its list, and the
implant's raw TLS connection to 193.57.9.155:4912 goes nowhere. The
public Ethereum RPC gateways the implant uses to find its server aren't
on that list either. Each refused connection is a red row in the
Security Timeline.
SafeDep also found that evjawbreaker checks for a debugger or a virtual machine and exits before connecting if it finds one. A Bromure machine is a virtual machine, and the layers above don't depend on the implant noticing.
Rebuilding the machine takes a click
The advisory's last line says to rebuild the machine from a trusted image. In Bromure you do that from a menu.
The backdoor lives in the home folder: ~/.solidity and
~/.config/systemd/user. A Bromure machine's home is its own disk image,
and Bromure takes a copy-on-write checkpoint of it every time the machine
boots. It keeps the last few boots, one a day for a week and one a week
for a month. Shut the machine down, open ⋯ › Rewind home…, and pick
the boot before November 9. The next start comes up with the home as it
was then, without the hidden folder and without the user unit. The current
home is kept as a new point in the same list, so you can go back to it if
you need a file from it.
Reset disk does the same for the system disk: it re-clones it from the
base image, and the home is untouched. A throwaway machine started with
bromure-cli vm run --rm is deleted when it stops. If the compromise alert
fired and you chose Shut down, the machine is marked compromised and
won't boot again until you wipe it, which keeps your settings and
credentials and replaces the disk and home.
Then the rotation step. The real tokens never entered the VM, so there's nothing there to rotate. If you want to rotate anyway, the real values sit in one place, the machine's Credentials pane on the Mac. Paste a new token there, and the stand-in in the VM stays the same, so nothing inside the machine needs editing.
Finding out who installed it
The first item on the advisory's list is a search: lockfiles, caches,
mirrors, CI logs, indirect copies. Bromure records every package fetch in
the Security Timeline, whatever the policy, with the package and
version in each row, for 90 days. Pick Supply chain in the engine
filter and you see npm · @ethersphere/[email protected] with a blue
checked — versions newer than the cooldown hidden row next to it,
which is the age gate keeping 13.3.0 out. Export… saves it as a CSV.
On Macs enrolled with bromure.io, each fetch is also sent to your
organization as a supply_chain.fetch event with the ecosystem, package,
version and outcome. "Did any agent in the company install bee-js 13.3.0
on November 9?" is one query, answered across every machine at once.
Bromure's traces then tie a download to the conversation behind it: the
tool call and the shell command the agent ran, and the model's turn that
asked for it. That's how you tell an agent that went looking for
swarm-cli from a build that pulled it in three levels down.
For the maintainers
The bee-js attack started with an unsigned bot commit that turned a
prepare script into a postinstall and added a network fetch to a
version-numbering script. It sat on GitHub for six minutes before the
release. A Bromure code review watches a repository with an agent on
one of your machines, on Every new commit or Every new pull
request. A reviewer that reads a diff like that one reports it.
The attackers also held a maintainer's GitHub account and the release bot's credentials, and nobody has said how they got them. If the agent on a maintainer's machine works in Bromure, its GitHub token is a stand-in, and new machines put every forge in Prompt before write, so a push or a release made from the VM waits for a dialog on the Mac.
Too young to install
The age gate ships on, at two days. npm in the machine never saw a version that was gone 75 minutes after release.
No hook to run
npm in the VM runs install scripts only for the agent CLIs. The proxy can also strip the hooks from every npm tarball.
Nothing worth taking
No wallet in the VM, only stand-in tokens, and a stand-in sent to the wrong host pauses the machine.
Rewind past it
Home checkpoints at every boot. Rewind home removes the backdoor, and Reset disk restores the system disk.
What to do this week
If you or your agents work with Swarm:
- Search lockfiles and npm caches for
@ethersphere/[email protected]and pin 13.2.0. - Check
bee-desktop.logforBee Desktop version: 0.56.2 (production)and move to 0.56.3. - Look for
~/.solidity/,solidity-langsupport-runner.serviceandcom.solidity.langsupport.runner, and hunt DNS and proxy logs forapi.frameworkmetrics.xyz,193.57.9.155and185.232.84.169. - If 13.3.0 or 0.56.2 ran, move the node's funds and stake to a new key from a clean computer.
If you publish packages, Socket's advice applies to any project: check in the publishing workflow that the commit is on a protected branch, require reviewers on the publish environment, and keep signing certificates away from workflows anyone can start on any branch.
Then give your agents a machine where a package like this one finds nothing to steal, and where you remove what it leaves behind from a menu. Setting the age gate, script stripping and egress rules is covered in the Supply-Chain Protection chapter of the manual.