npm Package Malware: Fake Web3 Accelerator Attack

Fake Web3 accelerator used npm package malware to hit crypto founders with a macOS infostealer and backdoor. Here is how the supply chain attack worked.

A young startup founder in a hoodie leaning over an open laptop at a cluttered kitchen table late at night, notebook and coffee mug beside him, warm lamp light in a small apartment

A VC Pitch That Installs a Backdoor

The artifacts are small and easy to miss in a lockfile: @apexacc/cli and @apexacc/audit, published to the npm registry by a user named apexaccelerator, linked to a GitHub account called apexaccs, and attributed to an entity that calls itself the Apex Foundation. Read the package descriptions and nothing looks wrong. One is an AI advisor for Web3 founders powered by Claude and Gemini. The other is a Model Context Protocol server exposing 18 tools for scanning tokens, wallets, and smart contracts.

That is the shape npm package malware has taken. The delivery mechanism is no longer a misspelled package name that catches a tired developer at 2am. It is a sales pitch.

The pitch went like this: free AI credits for Anthropic Claude and Google Gemini, a set of coding tools, and a due diligence scanner handed to crypto founders who are raising money. Elastic Security Labs, which published the research and tracks the operation as REF8801, describes tooling that “delivered information-stealing malware and backdoors for macOS and Windows machines,” per The Hacker News (opens in new tab). The people on the receiving end are not random. They hold wallet keys, deploy contracts, and run treasury operations from laptops that also carry their SSH keys and cloud credentials.

What makes this worth a close read is the inversion of trust. A founder who gets a warm intro to an accelerator and is told to run the diligence tool before the call is not being careless by typing npx. They are doing what the conversation asked for. The attacker understood that a pitch deck beats a typosquat because nobody runs a pitch deck through npm audit. Think about the last vendor you onboarded because a friend vouched for them, then think about how much of that vetting happened on the call versus on the terminal.

I have installed packages from people I met at conferences without reading a single line of the source. Sitting here now, that habit reads worse than it did a week ago, and I run production infrastructure for a living.

Two packages, one fake foundation, and a chain that terminates in a backdoor named CloudSyncD. The specifics are worth walking through, because the same pattern is already being repeated against MCP tooling that developers install by the dozen.

Background: How the Apex Foundation Manufactured Legitimacy

Start with the paper layer. Internet Archive snapshots from March 19, 2026 preserve the Apex Foundation’s own copy: it “offers $235,000 investments, grants, and partnerships, and connects projects with VCs, angels, and private equity if they don’t meet our criteria.” Read that sentence again as a founder would. The threshold to qualify is low, the upside is named in dollars, and there is a fallback offer of investor introductions if the money does not land. Nobody has to prove much to get a call.

Then there is the part that cannot be faked with a landing page. Elastic reports the group turned up at TOKEN2049 in Singapore and hosted smaller events of its own, which is credited as an attempt to borrow credibility from the room. A badge and a booth photograph do more for a founder’s instinct than any amount of website copy, because conferences are expensive and scammers are assumed to be lazy. The cost of a ticket is trivial next to the return on a single wallet with a healthy balance.

The package history ties the current campaign to an older one. In late July 2026, SafeDep documented three packages from the same operation shipping a macOS stealer from the Atomic Stealer family: @apexfdn/apex, @copilot-mcp/apex (with the same description later used on @apexacc/cli), and @apexfdn/copilot-mcp, billed as an Apex Copilot and MCP server for portfolio diligence. Three months between waves, same naming convention, same pitch.

The thread that actually connects them is infrastructure, not code. Socket captured artifacts linking the Apex Foundation to two domains since April 2026, apexaccs[.]org and apexfdn[.]xyz. Package names get burned and rebuilt; domains and the entity behind them persist. If you are ever doing triage on something like this, pivot on the resolution path and the GitHub account (apexaccs) rather than the npm scope, because the scope is the disposable part.

There is a useful analogy for anyone who runs a small hosting business. When a reseller account shows up with a plausible brand, a payment on time, and a referral from someone you know, you provision it and move on. You do not audit the business plan. This operation sold exactly that experience to founders, except the reseller account came with a launch daemon.

What’s Happening Now: Inside the npm Package Malware Chain

What were the packages advertised as? @apexacc/cli was sold as an AI advisor for Web3 founders, powered by something called “Claude Fable 5” and Gemini 2.5 Flash. @apexacc/audit claimed to be a Model Context Protocol server exposing 18 tools for scanning tokens, wallets, and smart contracts. Both names were published by an npm user called apexaccelerator and pointed back to a GitHub account named apexaccs. The descriptions read like a product page written by someone who had actually used the tools they were describing, which is the whole trick.

What did they install? On macOS, an encrypted loader that drops a backdoor tracked as CloudSyncD, disguised as an Apple component. On Windows, a custom infostealer that ships stolen data out through Telegram. Neither binary lives inside the tarball you download from the registry. The packages fetch platform-specific executables from the projects’ own GitHub releases, which keeps the published npm payload small and boring enough to survive a casual look.

How did the package decide which payload to serve? Both send an HTTP POST to vrf.apexaccs[.]org/api/v1/vrf with win32 or darwin in the request body. The response tells the client which stage to pull next for that OS. A single endpoint, two branches, and the server can change its mind per request whenever it wants, which is also handy for cutting off researchers who replay the call.

Did any of the promised functionality work? No. The 18 audit tools generate fabricated findings. The scoring test in the CLI returns an assessment of a project that does not exist. I have seen stub APIs before, but this is a different category. The tool is not unfinished, it is a prop. It exists so the first output looks like a real report and the founder pastes it into a Slack channel with the other due diligence notes.

The part that deserves attention is how ordinary the install instructions were. There was no curl pipe to bash in the README, no unsigned binary warning to click through. You ran the package, the package talked to a server, and the server decided what you got.

What npm Package Malware Like This Means in Practice

An MCP server package is not a library you import and forget about. It is a process that runs on your laptop or your builder’s laptop, with the same filesystem access, the same SSH keys, the same ~/.aws/credentials, and the same browser session you used to log into your exchange. Installing one because the setup instructions were three lines long is the same decision as clicking “Allow” on a browser extension that asks to read and change data on all sites, except there is no permission prompt at all. npm install @apexacc/audit is the whole consent dialog.

The Windows chain is worth memorising if you do endpoint triage. It hands off to a Base64-encoded PowerShell command that calls 0auth-chromium[.]com/APIService, which pulls drop.ps1, which fetches bludness.exe from ehegisnightmare[.]icu. That first domain is dressed up to look like a Chromium auth endpoint. It is not one, and if you see a PowerShell parent process reaching for it, you have a machine to isolate, not a browser to update.

The quietest part of the whole operation is the repackaged coding agent. Model requests get routed through attacker-controlled services, so your prompts and every file you paste into the context window travel somewhere else. Nothing crashes. The tool answers, the answers look plausible, and the only symptom is that your proprietary contract or your seed phrase recovery note now lives on someone else’s server. I have seen compromised developer machines with this shape before, and detection is miserable because there is no failure to alert on.

On attribution, Elastic spotted Russian-language comments in the commit history. I would treat that as a lead rather than a conclusion. Commit messages are trivial to write in any language, and a group that manufactures conference badges and fake VC pages is not above decorating its git log.

What to Expect Next

Copycats are the immediate worry. The barrier to entry here is a domain registration, an investment page with a dollar figure printed on it, and a conference badge that costs less than the monthly AWS bill for a mid-sized SaaS product. Nobody in this operation wrote a kernel exploit. They wrote a CLI wrapper, a fake MCP server, and a lot of theatre. That cost structure is going to get copied, and I expect the next round to aim at whichever vertical currently has cash and no patience for procurement. Biotech, legal tech, anything where founders are used to warm intros and moving fast.

The delivery vehicle will shift further toward agent and MCP tooling. These packages ask for broad filesystem and network access because that is the only way the feature set works, and users have spent the last two years being trained to grant exactly that without argument. A postinstall script at least has a known smell to it. An MCP server that reads your repo and calls an external model endpoint looks like the product doing its job. Free credits for a recognised model brand is also a far better lure than a cracked license key, because the offer is plausible in a way that pirated software usually is not.

Package removal is not containment. The npm packages and the GitHub repositories were already gone when Elastic published, but vrf.apexaccs[.]org, apexfdn[.]xyz, 0auth-chromium[.]com and ehegisnightmare[.]icu live in different registrars and different jurisdictions from the registry. Blocking the @apexacc scope in your proxy only helps until the next scope appears, which is a weekend of work for whoever is behind this.

I do not know how far the conference strategy scales, and I would be suspicious of anyone who claims to. But it only takes a handful of founders per event to make the economics work, and the trust transfer is badly asymmetric. One accepted coffee chat at TOKEN2049 buys a pass for every artifact that organisation ships afterwards. You cannot patch that, and no scanner is going to flag a business card.

Check Your Lockfiles Today

The scopes to hunt for are @apexacc and @apexfdn, plus the older @copilot-mcp. Run this across every repo you own, including the ones you think nobody runs anymore:

grep -rn "@apexacc\|@apexfdn\|@copilot-mcp" \
  package-lock.json yarn.lock pnpm-lock.yaml

Then on the machines themselves, because a lockfile only tells you what a project declared, not what someone installed globally while following a setup guide at 11pm:

npm ls -g --depth=0

If you are using pnpm, pnpm list -g --depth=0 covers the same ground. Bun keeps global installs under ~/.bun/install/global, so check there too if that is in your stack.

The macOS persistence check matters as much as the package check. Look for a launch item named com.apple.system.loader:

ls -la ~/Library/LaunchAgents /Library/LaunchDaemons

Apple does not ship anything by that name. I have seen plenty of legitimate com.apple.* labels in LaunchAgents belonging to third party apps that just borrowed the prefix, so do not panic at every match. Read the ProgramArguments in the plist and see what binary it points at. A plist calling something out of ~/Library/Caches or /tmp is the one you care about. Elastic’s writeup describes the loader as encrypted and disguised, and the disguise is precisely that label.

This is the same shape of problem as a compromised plugin on a shared host. The plugin folder is not the blast radius. Everything the web user could read is the blast radius, which is why restoring from a backup does not help if you restore to the same credentials.

If either package was installed on a machine, treat every wallet seed, every SSH private key in ~/.ssh, every AWS credential in ~/.aws, every browser session cookie and every saved password in that browser profile as exposed. Rotate them from a different device, not from the suspect one, because an infostealer that got in through a Windows chain that patched AMSI and unhooked ntdll.dll is not going to be politely observable while you reset things. Start with the wallets, since on-chain transfers are the one category of damage you cannot reverse with a support ticket.

So do the grep now, before you close the tab and go back to shipping.