n4uu

A DPRK RAT's changelog, read off the blockchain

· 11 min read

Seven TRON addresses pointed PolinRider loaders to 190 BSC transactions between June 2025 and October 2026, all sent by one wallet. 187 decrypt with the published keys. 95 are readable builds of the DEV#POPPER RAT, which grows from 7 commands to 15 over eight months before switching obfuscator. The archive also holds a Python port of the RAT published once in March 2026, three server addresses I could not find in public reporting, and the operator’s tests, ending with a payload that is just the string 'error'.

Contents

Summary

The loaders that PolinRider plants in GitHub repositories don’t carry their next stage. They read a TRON address’s latest transaction, take a BNB Smart Chain (BSC) transaction hash out of its memo, and decrypt that transaction’s input. The operator ships a new build by sending a new transaction, and nothing is ever deleted. So the chain holds every build that was ever served, in order, with timestamps.

I walked the full history of the seven TRON index addresses I could tie to the campaign. They hold 200 transactions from 6 June 2025 to 7 October 2026. 190 of them point to BSC transactions, all sent by the same wallet, 0x9bc1355344b54dedf3e44296916ed15653844509, to the burn address. 187 decrypt with keys that are already public, into 176 distinct scripts. 95 of those are builds of the DEV#POPPER RAT that eSentire described in March (89 distinct files), and 22 more are later builds in a different obfuscator. Lining them up gives the RAT a changelog:

Payloads published on chain by TRON index address, June 2025 to October 2026, and RAT commands per build rising from 7 to 15

This infrastructure is not new. Ransom-ISAC’s XCTDH series and Crystal Intelligence traced the BSC wallet, its funding and the RAT’s C2 addresses through November 2025, and eSentire, SafeDep, Socket, Greenflagged and Trend Micro documented individual builds. What I could not find in public reporting is:

How it was analysed

Nothing was run, and nothing talked to the operator’s servers. Every request went to public blockchain APIs: TronGrid for TRON and a public BSC RPC node. The decoding is the one Ransom-ISAC, SafeDep and StepSecurity published. Take the TRON memo, hex-decode it and reverse it to get a BSC transaction hash. Take that transaction’s input, hex-decode it, split it on ?.? and XOR the middle part with the loader’s key. Four keys from earlier reports cover all but three transactions.

The decrypted scripts went through crumble, which resolves javascript-obfuscator’s string array statically and can emulate a script against Node.js stubs that only record what it calls. I used the emulation only on the small loaders, to see which address or server each one reads next.

One check that the pipeline is right: BSC transaction 0x804b000af7d7e4337ba5db28bb367da64a08391de09ffb07847ac897c5f82954, published on 27 January 2026, decrypts to a file with SHA-256 207b5feb80c67f02023a07c7271b48c1d8b07ea5b59d76f98c1b0b357fbabfb1. That is the hash eSentire lists as “DEV#POPPER RAT (Obfuscated)”, and crumble’s output has the same thirteen commands as their table.

The archive

TRON indexWhat it servesTransactionsActive
TMfKQEd7TJJa5xNZJZ2Lep838vrzrs7mAPchain 1 loader; the 2025 builds read TLmj13…27Jun 2025 – Jun 2026
TXfxHUet9pJVU1BgVkBAbrES4YUc1nGzcGchain 2 loader, which fetches /$/boot from a C29Jun 2025 – Jun 2026
TLmj13VL4p6NQ7jpxz8d9uYY6FUKCYatSeRAT builds, first stream58Jun 2025 – Jan 2026
TA48dct6rFW8BXsiLAtjFaVFoSuryMjD3vRAT builds, second stream62Nov 2025 – Jun 2026
TCqf6ZkaQD84vYsC2cuu1jRwB6JveTaRrF2026 loaders, chain 129Nov 2025 – Oct 2026
TFMryB9m6d4kBMRjEVyFRbqKSV1cV2NcpH2026 loaders, chain 214Nov 2025 – Oct 2026
TPDBqDKAYfRVvVYDNZVxY8CRfovkwDaSgWthe Python RAT, once110 Mar 2026

Every BSC payload transaction, all 190, goes from 0x9bc1…4509 to 0x000000000000000000000000000000000000dEaD. That makes this the original EtherHiding transaction shape with one more hop in front, and it means a single sender address saw every build. Ransom-ISAC found that wallet’s first malicious transaction on 7 February 2025, and OpenSourceMalware found the same wallet behind the hijacked itsa-* npm packages of May 2025. Anyone watching that address would have seen each new RAT build as it was published.

The Aptos fallback addresses in the loaders are real accounts. They have sent 22, 10, 63 and 65 transactions, close to the counts of the TRON addresses they back up. The public Aptos fullnode no longer serves that history, so this post covers TRON and BSC only.

The changelog

The RAT is javascript-obfuscator output with an RC4 string array, and crumble reads 95 of the builds up to 18 March 2026. Command names are the ones eSentire documented. The date is the first build that carries each change:

First published (UTC)Change
12 Jun 2025First build, 31 KB, 7 commands: ss_info, ss_ip, ss_upf, ss_upd, ss_dir, ss_fcd, ss_stop. C2 136.0.9[.]8, 166.88.4[.]2
17 Jun 2025Size doubles to 62 KB. +ss_eval. First persistence marker, /*C250617A*/
19 Jun 2025+ss_inz, which writes the stager into other applications’ JavaScript files
20 Jun 2025C2 23.27.202[.]27 added
19 Aug 2025+ss_inzx, which removes it again
1 Oct 2025Second marker, /*C250618A*/
30 Oct 2025C2 202.155.8[.]173 added
14 Nov 2025Second stream starts on TA48d…, with its own key and only 198.105.127[.]210
19 Nov 2025+ss_connect, which switches C2 at runtime
16 Dec 2025+ss_eval64, +ss_cb (clipboard). Third marker, /*C250619A*/, the next day
30 Jan 2026Two builds carry a Python one-liner that runs whatever hxxp://154.91.0[.]196:443/download/2 returns
11–12 Feb 2026+ss_exit, +ss_exit_f. From here on no C2 address appears in clear text
25 Feb 2026Fourth marker, /*C250620A*/
18 Mar 2026Last build crumble can read: 75 KB, 15 commands
19 May 2026The RAT stream switches obfuscator. 22 more builds up to 8 June

The markers are already known: Trend Micro lists them and Ransom-ISAC’s YARA rule keys on C250617A. What the chain adds is dates. The first one, C250617A, shipped on 17 June 2025, which reads like a build date. The next three only count up, and they arrived months apart. Each new marker is kept alongside the older ones, presumably so that a new build recognises files that an old one already injected.

From November 2025 to January 2026 two RAT streams ran side by side. TLmj13… kept serving builds with the old C2 list, and from 1 December it added 198.105.127[.]210. TA48d… served only 198.105.127[.]210. The build eSentire analysed was the last one on the old stream.

From 19 May the RAT stream uses a different obfuscator: no javascript-obfuscator string array, generator functions and its own string decoding. Ransom-ISAC describes a new obfuscator in their September 2026 samples; I haven’t checked whether it’s the same one. crumble doesn’t handle it yet, so after 19 May I have sizes and dates but not contents.

The operator at the keyboard

The first transactions on TMfKQ… are not payloads. On 6 June 2025 the address received 100 TRX from the fee wallet Ransom-ISAC documented, TQdwohPCWqqfCUaCispyV1NaUZ1HgiJPUy. It then sent three 1-sun transfers with no memo, then one with the memo Mxy custom memo text, then the same memo to the TRON zero address. Someone was finding out how memos work.

On 11 June it tried putting the code itself in a memo. Two transactions, one minute apart, carry 1,024 and 8,192 bytes of a javascript-obfuscator script encrypted with cA]2!+37v,-szeU}. Both are cut off mid-script. The next day, 12 June, the first TRON memo holding a reversed BSC hash appears, on TLmj13…, and that is the scheme the campaign has used since. The same key, unused in between as far as the chain shows, fronted the second RAT stream from November.

On 25 February 2026, at 18:05 and 18:07 UTC, the same RAT build was published twice on TA48d…, first under cA]2!+37v,-szeU} and then under 2[gWfGj;<:-93Z^C. Everything after that uses the second key. That was a busy evening: 12 RAT builds and 13 loaders in under four hours, one of them with global["_V"]="00".

Test markers left in published loaders: _V='9-test' on 1 October 2025, _V="00" on 25 February, _V='A-00' on 23 May and global.i='00-test' on 5 October 2026. That loader and its chain 2 twin, published in the same minute, are the only ones that point to 89.19.222[.]25, at :443/init and /0/body.

The TRON side ends on 7 October 2026. That afternoon the chain 2 index, TFMry…, got three payloads of 13 or 14 bytes in 13 minutes. The last one, decrypted with that chain’s own key, is the JavaScript literal 'error'. A loader still resolving chain 2 now evaluates a string and stops there.

A Python port on a one-shot address

On 9 March 2026 the fee wallet TQdwoh… sent 100 TRX to a fresh address, TPDBqDKAYfRVvVYDNZVxY8CRfovkwDaSgW. The next afternoon that address made its only transaction, a pointer to a 10.5 KB payload that isn’t JavaScript. It’s a Python RAT built on asyncio, with the same command vocabulary: ss_shell, ss_eval, ss_eval64, ss_cb, ss_upf, ss_upd, ss_stop, ss_type, ss_info and ss_exit. Its configuration falls back to SERVER_HOST=154.91.0[.]15, SERVER_PORT=443 and an upload server at hxxp://198.105.127[.]210, where files go to /u/f. That is the upload path OpenSourceMalware saw in the 2025 itsa-* chain.

154.91.0[.]0/24 keeps coming back: .103 was the 2025 bootstrap server, .196 serves the Python stage in two January 2026 RAT builds and .15 is the Python port’s C2.

The second funder

Ransom-ISAC traced the fees for the first index addresses to TQdwoh…. The 2026 addresses were paid from somewhere else. TGJ1YJJNkFYLZKU4ZGqJfS6dU2cvGac7pS sent 50 TRX each to TFMry…, TCqf6… and TA48d… at 15:02, 15:03 and 15:04 UTC on 13 November 2025. That created all three addresses in one sitting, a day before TA48d… published its first build. The same wallet topped up TLmj13… on 30 October, three minutes before that address published two builds, and sent TA48d… another 100 TRX on 3 March 2026. TLmj13… was also funded by TQdwoh… in June 2025, so both fee wallets serve the same set of addresses.

Why it matters

Blocking a C2 address does little here, because the chain decides where the loader goes next. The chain is also the best telemetry anyone has on this campaign. For defenders that means three things:

Indicators

IndicatorRoleStatus
0x9bc1355344b54dedf3e44296916ed15653844509 (BSC)writes every payloadknown (Ransom-ISAC, SafeDep)
TGJ1YJJNkFYLZKU4ZGqJfS6dU2cvGac7pS (TRON)funded TFMry…, TCqf6…, TA48d…, TLmj13…not found in public reporting
TPDBqDKAYfRVvVYDNZVxY8CRfovkwDaSgW (TRON)one-shot index, Python RATnot found in public reporting
0x3414a658f13b652f24301e986f9e0079ef506992472c1d5224180340d8105837 (Aptos)fallback for TLmj13…not found in public reporting
hxxp://154.91.0[.]196:443/download/2Python stage fetched by two Jan 2026 RAT buildsnot found in public reporting
154.91.0[.]15:443Python RAT C2not found in public reporting
89.19.222[.]25:443/init, hxxp://89.19.222[.]25/0/body00-test loader, Oct 2026not found in public reporting
hxxp://198.105.127[.]210/u/fPython RAT uploadsaddress known (Ransom-ISAC)
4f7490cc2c0278f74097fb934d55dfaabfb6239e2fe27549895dc856b4d7a25bPython RAT

The seven TRON indexes are in the table above. Every payload hash with its date, index and BSC transaction, plus the network indicators, is in the CSV at the end of the post.

What the samples did to my tooling

On the first pass crumble resolved 216 string lookups in a RAT whose array has 1,467 entries. Two problems in the deobfuscator, both fixed:

  1. Wrappers that take their offsets from constant objects. This preset routes every string lookup through small functions like function aM(a, b) { return a0d(a - -a0d8.a, b); }, called as aM(a0db.a, 'zDx6'), with const a0db = {a: 479, d: 'Lr**'}. crumble already folded constant objects whose values were strings, but gave up on any object with a number in it, so neither the wrapper nor the call ever looked constant. Numbers now fold too.
  2. Folding a string concatenation can orphan a parenthesis. goja keeps no node for parentheses, so the span of 'a' + ('b' + 'c') ends at the last literal, and replacing it left the ) behind. The next parse failed, which silently ended the pass loop. The RC4 keys made this worse, because some of them contain parentheses ('5)#)', ')J^Z'). The fix widens the span until its parentheses balance, skipping quoted text.

With both, the same file resolves 1,491 calls with no failures, and every command eSentire lists is in the output.

References

Cite

n4uu, "A DPRK RAT's changelog, read off the blockchain", 17 November 2026. https://n4uu.github.io/2026/polinrider-rat-changelog-on-chain/

Indicators as CSV: polinrider-rat-changelog-on-chain.csv (defanged).