Contents

A Recruiter Sent Me a Take-Home Test. It Takes Its Orders From an Ethereum Wallet.

A fake recruiter, a VS Code extension that shows one script and runs another, and malware that looks up its server in Ethereum transactions. Read without running a line of it.

On 10 November a recruiter emailed me about a DevOps Engineer role. The email came to the alias from my post in the October “Who wants to be hired?” thread on Hacker News and named two companies from my CV. The next step was a coding assessment inside VS Code, through an extension from the official Marketplace.

I didn’t install it. I downloaded the package, read it, and followed every address it pointed to. At the end of the trail there is a remote-access tool and a credential stealer. The part I enjoyed is in the middle: the malware finds its server by reading the latest transaction of an Ethereum wallet. That wallet is public, so four months of the operator’s server moves, working hours and money are public too.

From inbox to remote shell in six steps
Pick a step. Everything below was read from files and public records; none of it was run.
A recruiter persona offers a remote DevOps role and a take-home test that runs inside VS Code. The message quotes my public work history.
Sender domain registered on 3 November, seven days before the email
Namecheap private mail, DMARC policy set to none
iCloud had already flagged it as spam

The email

The message itself was fine: the right role, real employers (both are on my public profile). The problems were one layer down. The sender’s domain, huntrole.org, was registered seven days earlier, the DKIM signature didn’t verify, and iCloud had already filed the message as spam. And the assessment could only be done in an extension that starts together with the editor.

The careers site won’t take an application without a PDF resume, so it collects CVs even from people who never open VS Code. I sent it an invented candidate and got a ticket ID and instructions pointing at the extension.

The extension shows one script and runs another

There is no malware in the package, and that’s the design. On every editor start the extension (huntrole.assessment-tool, 126 installs on 16 November) sends the hostname and VS Code’s machineId to api.huntrole.com and keeps a WebSocket open.

When the assessment panel opens, the server returns a task with two code fields. The panel “types” displayScript on screen, then passes script to new Function with the real require and process: files, network, child processes. No ticket and no Run button needed.

One task, two scripts
The server decides what you read and what your machine runs, separately. Switch the device state to see what changes.
displayScript typed out in the panel for you to read
const search = (xs, target) => {
let lo = 0;
let hi = xs.length - 1;
while (lo <= hi) {
const mid = (lo + hi) >> 1;
if (xs[mid] === target) return mid;
if (xs[mid] < target) lo = mid + 1;
else hi = mid - 1;
}
return -1;
};
const result = search([1, 3, 5, 8, 13], 8);
script passed to new Function() with require and process
const search = (xs, target) => {
let lo = 0;
let hi = xs.length - 1;
while (lo <= hi) {
const mid = (lo + hi) >> 1;
if (xs[mid] === target) return mid;
if (xs[mid] < target) lo = mid + 1;
else hi = mid - 1;
}
return -1;
};
const result = search([1, 3, 5, 8, 13], 8);
console.log("result:", result);
One more line follows: an obfuscated loader, about 8.3 kB. The text above is placeholder noise. The real code is not reproduced here.
Size of the executed field 150 bytes
A harmless example and nothing else. The displayed field is the same text without the last line.
Sizes are from two snapshots of the same endpoint, taken minutes apart with a synthetic device. The readable code is my stand-in for a benign task from a third capture.

For a device the server didn’t know, script was 150 harmless bytes. After my invented device registered and opened the socket it was 8,665 bytes, and displayScript hadn’t changed. So a scanner sees a clean package, and so does a candidate who reads the files before installing.

The server address is a transaction

The long script has no server address in it. It has a wallet address and code that looks up the wallet’s latest transaction. With the names cleaned up:

const wallet = "0x1eEf0849EDa15d87f64e51C73C8442Cf3c8bCEF3".toLowerCase();

function decode(to) {
  const bytes = Buffer.from(to.replace(/^0x/i, ""), "hex");
  const ip = (b) => b[0] + "." + b[1] + "." + b[2] + "." + b[3];
  return [ip(bytes.subarray(0, 4)), ip(bytes.subarray(4, 8))];
}

to is the recipient of that transaction. An Ethereum address is 20 bytes, and the operator fills them with an IPv4 address, the same address again, and 12 bytes that spell helloipbot!!. The transaction moves zero ETH and nobody owns the recipient.

From here on, that recipient is a marker, a transaction to it is a ping, and the sender is a pointer wallet.

A server address dressed up as a wallet address
Every transaction the pointer wallet sends goes to an address like this. Pick one that was seen on chain, or type an IPv4 address and forge your own.
Evoxt, AS149440 Virtuo, AS399486 not routable, test markers
0x5f87b5545f87b55468656c6c6f6970626f742121
5f 95
87 135
b5 181
54 84
bytes 1-4, read as an IPv4 address
5f 95
87 135
b5 181
54 84
bytes 5-8, a second address
68 h
65 e
6c l
6c l
6f o
69 i
70 p
62 b
6f o
74 t
21 !
21 !
bytes 9-20, plain ASCII, the same in all 13 markers
95.135.181.84 port 443 for the remote-access tool, port 80 for the staged payloads
Virtuo, AS399486. 19 pings from the HuntRole wallet, 12 Nov, 23:44 to 15 Nov, 11:07 UTC.
One request from your browser to the public index eth.blockscout.com. Nothing is requested from the operator's servers.
Markers decoded from 1,157 transactions of three wallets, as of 2026-11-15. In the eleven live markers both halves are equal; they differ only in the two test markers the operator wrote on the first day.

The script doesn’t search for the ping. It rounds the current block number down to a multiple of 1,000 and opens six blocks around the round numbers, so the writer has to be punctual. If those are empty it falls back to a block explorer, and nothing checks the record’s age, so a wallet that has stopped writing still points at its last server.

For the operator this is cheap and hard to take down: no registrar can remove the pointer, and a ping costs about 0.00007 ETH. In exchange, every server they ever used, and the minute they switched to it, stays on a public ledger for good.

Three wallets, eleven servers

Every marker ends in the same 12 bytes, so the wallets can be found from the other side. I scanned every Ethereum transaction since 1 April for that ending and got three senders. Between them: 1,157 pings to 11 servers from July to 15 November.

Four months of pings, three wallets, thirteen markers
Each bar is an unbroken run of pings to one server address. Hover a bar to see the address; bars for the same server light up together across wallets.
Evoxt, AS149440 Virtuo, AS399486 not routable, test top-up from the treasury top-up from elsewhere
Jul Aug Sept Oct Nov 2026, UTC 13-16 Oct, linked rotation 10 Nov, the email arrives Public reports Socket NullReceiver named SafeDep GitGuardian Sonatype Malicious packages fetch-page-assets two PyPI packages HuntRole front huntrole.com registered extension published Early wallet 0xee36...4527 .62 Second wallet 0xa322...Ef1a .62 .135 .46 .38 .75 .174 .38 .155 HuntRole wallet 0x1eEf...CEF3 .75 .151 .84
Show the runs as a table
WalletAddress in the markerNetworkFirst ping, UTCLast ping, UTCPings
Early wallet 252.21.32.45 not routable 12 Jul, 10:19 12 Jul, 10:59 3
Early wallet 100.100.121.22 not routable 12 Jul, 11:19 14 Jul, 06:36 13
Early wallet 166.88.134.62 Evoxt AS149440 14 Jul, 06:46 17 Jul, 05:05 220
Early wallet 100.100.121.22 not routable 28 Jul, 07:35 28 Jul, 07:35 1
Early wallet 166.88.134.62 Evoxt AS149440 29 Jul, 08:03 30 Jul, 01:34 9
Early wallet 100.100.121.22 not routable 31 Jul, 06:18 31 Jul, 06:18 1
Second wallet 166.88.134.62 Evoxt AS149440 25 Jul, 06:48 22 Aug, 00:08 178
Second wallet 23.27.13.135 Evoxt AS149440 22 Aug, 01:28 27 Aug, 23:57 48
Second wallet 166.88.73.46 Evoxt AS149440 28 Aug, 01:46 3 Sept, 13:13 61
Second wallet 193.247.144.38 Evoxt AS149440 4 Sept, 00:49 17 Sept, 08:24 85
Second wallet 166.88.134.75 Evoxt AS149440 17 Sept, 10:35 28 Sept, 09:16 92
Second wallet 91.218.183.174 Evoxt AS149440 28 Sept, 10:18 6 Oct, 04:38 67
Second wallet 193.247.144.38 Evoxt AS149440 6 Oct, 07:25 13 Oct, 23:57 44
Second wallet 23.27.13.72 Evoxt AS149440 14 Oct, 00:06 14 Oct, 00:14 5
Second wallet 103.179.142.155 Evoxt AS149440 16 Oct, 05:24 19 Oct, 10:22 27
HuntRole wallet 166.88.134.75 Evoxt AS149440 4 Sept, 05:18 12 Sept, 11:21 61
HuntRole wallet 103.179.143.219 Evoxt AS149440 14 Oct, 01:52 14 Oct, 03:18 13
HuntRole wallet 23.27.13.72 Evoxt AS149440 14 Oct, 04:52 14 Oct, 04:52 1
HuntRole wallet 103.179.143.219 Evoxt AS149440 14 Oct, 04:54 16 Oct, 02:07 15
HuntRole wallet 95.135.181.151 Virtuo AS399486 16 Oct, 05:00 12 Nov, 22:57 194
HuntRole wallet 95.135.181.84 Virtuo AS399486 12 Nov, 23:44 15 Nov, 11:07 19
1,157 zero-value transactions read from the public Ethereum history on 2026-11-15. A run breaks when a wallet stays silent for more than 18 hours. Package dates are the days the advisories were opened, not the upload dates.
  • The early wallet worked for three weeks in July and starts with a developer’s test transactions: the text hello as data, markers that decode to addresses no packet can reach. I couldn’t find it in any public report.
  • The second ran from 25 July to 19 October, with 607 pings to eight servers. It’s the wallet from the August reports on trojanised npm packages.
  • The third is the one in my script: 303 pings to five servers since 4 September. It last switched on 12 November, two days after the email.

Nine of the eleven servers, and api.huntrole.com too, sit at one hosting company, Evoxt. The two newest are at another, Virtuo, side by side in one /24.

The timer counts blocks

A timer sends the pings. A person pays for them.
Top: the gap between consecutive pings, one tick per gap, on a logarithmic scale. Bottom: the hour of day, in UTC, at which each wallet signs.
200.6 min
median gap between pings for both long-running wallets: 1,000 Ethereum blocks
0.3%
spread of that gap for the HuntRole wallet: 37.5 s over 270 cycles
31.6 days
longest silence, 12 Sep to 14 Oct, with gas for roughly a hundred more pings still in the wallet
1 min 10 min 1 h 1 day 10 days 10 blocks, 2 min 100 blocks, 20 min 1,000 blocks, 200.6 min Early wallet 246 gaps 11.1 days Second wallet 606 gaps 3.1 days HuntRole wallet 302 gaps 31.6 days
00 06 12 18 37
Early wallet
74 of 247 pings fall between 07:00 and 08:59.
00 06 12 18 49
Second wallet
607 pings, every hour of the day is used.
00 06 12 18 21
HuntRole wallet
303 pings, no preferred hour.
00 06 12 18 80
Treasury
284 signed transactions on 14 chains, on 37 days. None between 12:00 and 22:59.
Each dial is scaled to its own busiest hour; the number in the centre is that hour's count. A timer that fires every 200.6 minutes drifts through the clock, so its dial fills evenly. The treasury's dial covers 37 days.

The typical gap between pings is 200.6 minutes, which is exactly 1,000 blocks. Of the third wallet’s 303 pings, 270 landed in a block whose number ends in 999, 000 or 001, the blocks the script opens first. The timer keeps a fresh copy of the record where the reader will look, even when the address hasn’t changed.

The treasury that pays for this keeps hours. Its key has signed 284 transactions on 14 chains over 37 days, none of them between 12:00 and 22:59 UTC. The email’s Date header says +0500, and on that clock the signing hours look like a working day. Both are trivial to fake, so I won’t name a time zone.

Where the gas money comes from

Where the money comes from and where it goes
Top: the operator's treasury, which mostly cashes out a token portfolio. Bottom: the pointer wallets, which run on pocket change. Hover a box to trace its transfers.
The money The pointer system Nine wallets on 3 Sep one funding history since 2021 27 other addresses two also pay the same people Treasury 0xf970...ecd0 LI.FI bridge and swaps 1inch swaps FixedFloat deposit addresses Other routers, small payments 0x96c8e62c fed by a 102-address disperser FixedFloat hot wallet Swap routed through LI.FI Service wallet, two relays Second wallet 607 pings HuntRole wallet 303 pings Early wallet 247 pings Gas for 1,157 pings 1.660.4692.470.1190.1140.03240.02700.06000.02930.03000.01760.02070.01150.01890.04010.02060.0190
Show the transfers as a table
FromToETHWhen, 2026
Nine wallets on 3 Sep Treasury 1.66 3 Sep
27 other addresses Treasury 0.469 4 Sep to 15 Nov
Treasury LI.FI bridge and swaps 2.47 3 Sep to 15 Nov
Treasury 1inch swaps 0.119 6 and 11 Nov
Treasury FixedFloat deposit addresses 0.114 28 Sep, 19 Oct
Treasury Other routers, small payments 0.0324 22 Oct to 1 Nov
Treasury Second wallet 0.0270 13 and 16 Oct
Treasury HuntRole wallet 0.0600 11 Nov
Second wallet Treasury 0.0293 14 Oct
0x96c8e62c Second wallet 0.0300 30 Jul
FixedFloat hot wallet Second wallet 0.0176 25 Jul
FixedFloat hot wallet HuntRole wallet 0.0207 14 Oct
Swap routed through LI.FI HuntRole wallet 0.0115 4 Sep
Service wallet, two relays Early wallet 0.0189 11 to 28 Jul
Second wallet Gas for 1,157 pings 0.0401 25 Jul to 19 Oct
HuntRole wallet Gas for 1,157 pings 0.0206 4 Sep to 15 Nov
Early wallet Gas for 1,157 pings 0.0190 11 to 31 Jul
Ethereum only, amounts in ETH, line width follows the square root of the amount. The dashed line is the one transfer that went back. Token transfers and swap proceeds are not drawn, so more ETH leaves the treasury than enters it here. The two boxes with a bold outline are the same exchange.

The pointer wallets have burned about 0.08 ETH in four months. One treasury wallet funded two of the three, and each of its three top-ups was followed by a server change: 7 minutes, 8 minutes and 44 hours later.

The treasury’s own money looks stolen, though the ledger alone can’t prove it. On 3 September nine wallets, idle for months and mostly tied to each other by old transfers, sent it what they held within half an hour, on Ethereum and several other chains. The treasury sold the tokens, bridged about 12,560 USDC to BNB Chain and cashed out through one-time addresses of what looks like a swap service. Nine days later someone, I think the owner, moved what was left to a fresh address, and none of it went to the treasury.

It happened again on 3 November. A private wallet sent the treasury 5,344 USDT on BNB Chain and a little ETH on Ethereum within four minutes, and about twenty minutes later the treasury swapped it and moved it on.

Wallet files are on the stealer’s list, but nothing on chain says it emptied these wallets. Their addresses stay out of this post.

The treasury also pays people: 32 USDT transfers to eight addresses since September, 3,813 USDT in total, at most 21% of what those addresses received this year. Two of the eight cash out through an address that forwards everything to a wallet the block explorer labels OKX, and two more into one labelled Lemon Cash. Both exchanges verify their customers. If the labels are right, someone with the authority to ask can get names.

Going in, there is no such door: two of the pointer wallets got their first ETH from FixedFloat, an exchange that asks for no identity.

The same wallets feed npm and PyPI

The recruiter front is one delivery channel of several. The second wallet’s address is quoted in at least 40 advisories in the OpenSSF malicious-packages repository, nearly all for npm. On 9 November JFrog described another fake recruiting company, AccuraSoftHire, whose take-home project reads the same wallet through the same paths with the same XOR keys. And the October reports on hijacked npm packages from GitGuardian, OX Security and Sonatype name the third wallet, the one in my script.

What arrives on the machine

I fetched each stage once from a disposable VM and decoded it without executing anything: two small loaders, a remote-access tool, a 1.4 MB injector that plants code into other applications, and a Python stealer.

What gets read and what gets left behind
Two different programs. The Python stage collects and uploads. The JavaScript tool is the one that changes files, and its changes outlive the extension.
Read and uploaded Python stealer, 92.7 kB decoded
process.env
every environment variable, saved as _sysenv.json and _sysenv.env: API keys, cloud credentials, tokens
Browser profiles
saved passwords, cookies and databases; Firefox profiles and extensions
OS secret stores
Windows DPAPI, Linux Secret Service and KWallet. The macOS Keychain query needs a flag the boot stage does not pass
~/Library/Keychains/login.keychain-db
copied as a file on macOS; a copy is not a decryption
~/.git-credentials
and ~/.config/git/credentials
~/.config/gh/hosts.yml
the GitHub CLI token
VS Code globalStorage
whatever your extensions keep there
Wallet files
Solana id.json, Exodus, Electrum, Ledger, Trezor
Password managers
1Password, Bitwarden, KeePass data files
Sent with POST /u/e and POST /u/f, optionally through Telegram if the server hands out a bot token.
Written and kept JavaScript remote-access tool
~/.node_modules/.npm-ws.json
configuration
~/.node_modules/.npm-ws.pid
process control
~/.node_modules/.npm-ws-runner.js
the loader itself
npm-cli.js, lib/cli.js
npm is wrapped, so the runner starts on later npm calls
*.inz.cjs
a loader dropped next to an Electron app's own code
*.inz.orig
backup of the file that was patched to load it
Electron apps on the list
VS Code, Cursor, Antigravity, Discord, GitHub Desktop; on macOS inside /Applications/.../Contents/Resources/app
Uninstalling the extension removes none of this.
This is what the code tries, read statically. Whether it succeeds on a given machine was not tested.

The remote-access tool is byte-identical to a published DEV#POPPER sample. It dials out, so a home router doesn’t stop it, and it accepts arbitrary shell commands. The stealer copies browser databases, keychain files and wallet files. Since I ran nothing, I can’t say what it actually gets from a real machine. The injector I mostly couldn’t read.

Who is behind it

Vendor reports on related incidents attribute this family (PolinRider, DEV#POPPER) to North Korean groups. My data ties HuntRole to the family through identical bytes, shared wallets and shared servers. It says nothing about the country.

Is this the crew that took about 6.3 billion dollars from Binance on 29 October? By Binance’s account that attack also went through a fake hiring process with a coding assignment. None of the 453 addresses around this ledger is among the 608 Ethereum addresses in the FBI alert, and Binance has published no hashes or servers to compare. I found no link and can’t rule one out.

If you installed it

Check for the extension and for what the remote-access tool leaves behind:

code --list-extensions | grep -i huntrole
ls -la ~/.node_modules/ 2>/dev/null | grep -i npm-ws
find ~ /Applications \( -name "*.inz.cjs" -o -name "*.inz.orig" -o -name "*.inz.tmp" \) 2>/dev/null
  • Installed, panel never opened. The server has your hostname and machine ID. As far as the code shows, nothing else ran. The extension can open the panel on its own, though, so if in doubt count it as opened.
  • Panel opened. Assume the script ran, whatever the checks show.
  • Any of those files exist. The machine is compromised. From another device, rotate everything your account could read: .env files, browser passwords and sessions, Git and cloud credentials, wallet seeds. Then reinstall Node and the patched applications, or rebuild the machine.

When Microsoft removes a malicious extension, VS Code uninstalls it automatically (so its documentation says). Whatever the tool planted stays.

Who has been told

On 15 November I reported the extension to Microsoft, the domains and mailbox to Namecheap, the careers site to Vercel, and the servers to Evoxt and Virtuo. Two hours later I had automatic acknowledgements. A day later the extension is still listed.

The operator was busy that day too. Five hours before my reports the treasury signed 29 transactions on ten chains, most of them moving its leftovers to an address it had never used. The next morning that address signed for the first time: it swapped some BNB into USDT and sent 60 USDT to one of the eight addresses the old treasury had been paying. It pays the same people, so I’m calling it the new treasury.

How I read it without running it

Coding agents did most of the reading, and a sample can talk to the model that reads it: a planted AGENTS.md is a prompt injection attempt. So the sample was only ever data, the model reading it had no network, no shell and no secrets, and anything dynamic ran in a disposable microVM with the network closed by default.

Indicators

Wallets, markers, servers, paths and hashes
# pointer wallets
0x1eEf0849EDa15d87f64e51C73C8442Cf3c8bCEF3   HuntRole script; October npm and GitHub wave
0xa322E5f3D311D3080e6f0121063e9aDC2490Ef1a   NullReceiver packages, 25 Jul to 19 Oct
0xee36cc64f82c0b891df975452b91fcbe1b194527   early wallet, 11 to 31 Jul

# treasury
0xf97062cb1b6ffa18b88e9faea373ed288a87ecd0
0x4e3e0a353dc7a5c24a133ab246de08a177f6cd4c   new treasury: took the leftovers on 15 Nov, first payout on 16 Nov

# markers: 0x + ip + ip + "helloipbot!!" (68656c6c6f6970626f742121)
0x5f87b5545f87b55468656c6c6f6970626f742121   95.135.181.84    current on 16 Nov
0x5f87b5975f87b59768656c6c6f6970626f742121   95.135.181.151
0x67b38fdb67b38fdb68656c6c6f6970626f742121   103.179.143.219
0x67b38e9b67b38e9b68656c6c6f6970626f742121   103.179.142.155
0x171b0d48171b0d4868656c6c6f6970626f742121   23.27.13.72
0x171b0d87171b0d8768656c6c6f6970626f742121   23.27.13.135
0xa658864ba658864b68656c6c6f6970626f742121   166.88.134.75
0xa658863ea658863e68656c6c6f6970626f742121   166.88.134.62
0xa658492ea658492e68656c6c6f6970626f742121   166.88.73.46
0xc1f79026c1f7902668656c6c6f6970626f742121   193.247.144.38
0x5bdab7ae5bdab7ae68656c6c6f6970626f742121   91.218.183.174

# test markers, not routable, halves differ
0xfc15202df10b150b68656c6c6f6970626f742121   252.21.32.45 / 241.11.21.11
0x646479162d0c1f0b68656c6c6f6970626f742121   100.100.121.22 / 45.12.31.11

# domains and hosts
huntrole.com        careers site, Vercel
jobs.huntrole.com   Vercel
huntrole.org        sender domain, registered 2026-11-03
api.huntrole.com    23.27.22.119, Evoxt
[email protected]   sender address

# extension
huntrole.assessment-tool 1.0.0
VSIX SHA-256 26761cc3bc24fde4aaf89fc0fc3cf8fbfb7bfb5e4961e85cf3c9f9e7732174bd

# extension API
/api/v1/careers/generate-rand-coding
/api/v1/devices/register  /api/v1/devices/bind-ticket  /api/v1/devices/verify-ticket
/socket/presence
headers: x-hostname  x-machine-id  x-local-candidate

# server paths
/0x/cls  /0x/ls  /0x/clb  /$/boot  /$/1  /u/e  /u/f  /verify-human/<channel>
header: X-Payload-B64

# chain is read through
eth.blockscout.com/api  1rpc.io/eth  eth.drpc.org
ethereum-rpc.publicnode.com  eth-mainnet.public.blastapi.io

# XOR keys
/0x/cls         q4FZkxX{!h,Sr3=@
/0x/ls          y-p_>d$0B&@^1aQk
loader strings  em0ZJl7uATsA

# persistence
~/.node_modules/.npm-ws.json
~/.node_modules/.npm-ws.pid
~/.node_modules/.npm-ws-runner.js
*.inz.cjs  *.inz.orig  *.inz.tmp

# SHA-256 after decoding
script (verified)  23bc5f13eacfd14eb7dbf370a147a4d040a1e95f6c7d49262dcf2eac36f50fc4
remote-access tool 8252068e04cad13c326fb7b833c4db8f45522a52984b7315c089b8b914babce1
/0x/cls            dc45f264c338c024b67c666cea25ffb247e0e4261cf40af6a4ddd2627ccc722f
/0x/ls             228f3c5924325161c398a6c2b66334397b850b6d9248fd1a6ccd65ba1b567eda
/$/boot            a4a9426cea9670d1acc19f263f874fb486c8650540392a517f2d431f0ce94609
/$/1 (python)      9e842de8444b8513cd93e654c913297e1a84237a21a52d511c12a96a824f44a5
injector           ff57ef4999470c3bce2224360f6e216b795e78d4eb8d0e1fb2ae8f9eff95cdb6

Sources