For years, the weakest link in a digital espionage campaign has been the stretch in which the implant asks the command server for permission to download the next piece. That stretch travels encrypted, but with algorithms that an analyst with time and hardware can attack. The North Korean group Lazarus swapped that piece for post-quantum cryptography, according to the analysis that BleepingComputer published on October 21, 2026: the LPEClient malware receives ML-KEM public keys from the server and uses them to decrypt the executable that completes the infection.
ML-KEM is no exotic invention. It is the key encapsulation mechanism that NIST approved in August 2024 as FIPS 203, the standardized version of CRYSTALS-Kyber, designed so that communications survive a quantum computer that has not yet been built.
The finding matters for two reasons that point in different directions. The first is that a group with state resources adopted a 2024 cryptographic standard before much of the commercial software that should protect its customers. The second is that Kaspersky does not attribute the move to a desire to comply with standards: in its reading, the goal is for subsequent payloads to travel out of reach of whoever is watching the network. Without the private key, moreover, intercepted traffic yields only unstructured bytes.
A fake job lure and a chain that lives in memory
The first evidence of FlashHole dates to July 2025, according to Kaspersky's technical report, and the pattern repeats that of the SyncHole operation, documented in April 2025. Kaspersky believes, with medium-high confidence, that the targets received emails with links to websites hosted on GitHub Pages. One of them was the résumé of an alleged network engineer; the other, an interior design catalog aimed at the construction sector. The candidate's photo had no EXIF metadata and its texture looked unnatural, which led researchers to attribute it to an image generator.
On opening those pages, a script loaded from an external domain triggered the entire chain in memory. The link that made it work was INISAFE CrossWeb EX, a financial security program mandatory in South Korea, in its version 3.4.1.2, published in 2023. The chain relies on software the user cannot remove without losing access to online banking or government procedures.
That legitimate process ended up executing fsquirt.exe, a Windows binary, and the shellcode that installed TechHub was injected into it; TechHub is the downloader that opens the operation and in this version uses private GitHub repositories with an access token.
Kaspersky identified at least three affected companies in the software, construction and wired communication equipment manufacturing sectors, and suspects the real figure is considerably higher. The sectors do not match those of SyncHole, a sign that the group has broadened its target.
Four 1,568-byte blocks and a Threefish key
LPEClient is the only piece of the arsenal that Kaspersky had already seen in SyncHole, and also the one that has changed most in this campaign. In the new version it lost the command server address and talks to it through ClientHub, a newly identified loader that in some samples uses Dropbox as a channel.
That move complicates blocking by domain name, and Kaspersky reads it as an attempt to gain concealment: with no fixed address to point to, the defender is left only with the encrypted content.
The first message carries the profile of the compromised machine compressed with XPRESS Huffman and encrypted with GOST, the Russian standard also known as Magma, with a random key that travels at the start of the packet. The response to that send is decrypted with AES-CBC-256, using as key the SHA-256 hash of a sequence of 1,568 random bytes that the implant had previously sent to the server.
That response brings four 1,568-byte blocks. Each contains an ML-KEM-1024 public key, and on them LPEClient generates four pairs of shared secret and ciphertext that return to the server encrypted with GOST. The four secrets, joined, form the 256-byte key with which the next message is decrypted using Threefish, a block cipher that accepts a tweak value in addition to the key.
| Step |
What LPEClient does |
| 1 |
Profiles the victim and sends the report encrypted with GOST |
| 2 |
Receives four blocks with ML-KEM-1024 public keys |
| 3 |
Derives four shared secrets and sends them with GOST |
| 4 |
Composes the 256-byte Threefish key |
| 5 |
Decrypts the MZ payload and executes it in memory |
The decrypted message has a strict format: first the length of the parameters, then the parameters and finally the payload, which must exceed 4 KB and begin with the characters MZ for the implant to treat it as a Windows executable. LPEClient loads it reflectively and the end result is Cubridge, the orchestrator that deploys modules, plugins and tools on the compromised machine.
Why would an attacker adopt post-quantum cryptography now?
The reasonable question is what a criminal group gains from an algorithm designed for a threat that has not yet arrived. It does not need a quantum computer to benefit from ML-KEM, because the standard solves the distribution of a secret key between two parties with a primitive that still stands after years of public scrutiny and that already ships implemented in open, fast libraries.
For banking and for infrastructure, the argument for these primitives is longevity. Data encrypted today with RSA or with elliptic curves can be stored and decrypted when the machine capable of doing so arrives, so the secret to be protected outlives the algorithm that covers it. In FlashHole the benefit is different, and Kaspersky says so bluntly when it states that the use of post-quantum cryptography shows the intention to conceal subsequent payloads.
The calendar is already being discussed in other sectors. Zcash wants to bring quantum-resistant signatures in January 2027, as we reported when covering the announcement, and Vitalik Buterin warned that AI could break crypto's security before quantum computing does.
. The difference is that in this case the one who migrated first was the attacker. Kaspersky closes its analysis with the irony that the group adopted these protections faster than many legitimate security products, and the reproach is aimed at software that has not yet taken that step.
What FlashHole reveals about Lazarus
Kaspersky places FlashHole within Cascade Hole activity and attributes it to the same hand as SyncHole with medium-high confidence, relying on the entry vector, the targets and the shared use of LPEClient. The group chains campaigns without pause: the oldest evidence of FlashHole is from July 2025, just as the previous operation was winding down after at least four months of activity.
The same ecosystem has an economic leg. Researchers saw COPPERHEDGE in parallel, with a case that ended in data theft to a leak site set up for the occasion and in an episode that AhnLab linked to the Gunra ransomware.
The dismantling has dates. In October 2025 the South Korean government presented a plan to remove the mandatory installation starting in 2026, and in July 2026 it published a joint advisory with several manufacturers on these techniques. Kaspersky expects operations against these tools to remain active as long as they exist, and a paper presented at the USENIX Security 2025 symposium had already measured the attack surface they generate. Software was the only sector present in both campaigns, a detail analysts read as interest in source code and in the vulnerabilities it opens up afterwards.
The bar the malware has just raised
FlashHole does not uncover a new flaw or debut an exploitation technique. What changes is the ground detection rests on: if the delivery channel uses modern algorithms, whoever intercepts traffic gets nothing readable and has to go after the memory of the infected machine, which is precisely where Lazarus leaves the least trace.
The useful question is not when the quantum computer will arrive, but how long the rest take to copy the idea. ML-KEM has been in FIPS 203 since 2024, it is free and it already travels inside a malware chain. Any defensive team that still treats command-channel encryption as a secondary nuisance will be late to the next campaign.