Seven RAT families, one actor: what SilkParasite does to behavioural clustering

Clustering works because engineers repeat themselves. Analysts join intrusions together on copied code, on a command and control protocol an author reuses because it already works, and on the build residue nobody bothers to change between jobs. SilkParasite's seven families sit across at least four language toolchains, .Net, C++, Go and JavaScript, with C2 run through Google Drive alongside HTTP, TCP and DNS ^[5][25]. Bitdefender adds that the code was deliberately built not to resemble previous families, with a minimum footprint and in-memory execution ^[23]. Each family needs a loader and a transport rather than a full feature set, because the capability arrives later as plugins for keylogging, clipboard monitoring or shell access ^[6].

The arithmetic in the report is worth doing. Five of the seven families were previously undocumented ^[1], which leaves two that were not ^[21]. The single thread running out of this campaign into another tracked cluster comes from one of those two: BloodAlchemy sits in the ShadowPad and Deed RAT lineage, and Deed RAT was FamousSparrow's primary backdoor ^[17]. None of the five that Bitdefender named for the first time, DriveSilkRAT, CookiETagRAT, NomadRAT, GoginRAT and NodeEdgeRAT ^[3], carries a comparable documented link ^[22]. What was written fresh stayed unattached. The handle came from inherited code.

That is a narrower problem than it first looks, because clustering in this ecosystem was already leaky. Bitdefender notes that tooling, techniques and infrastructure circulate across otherwise distinct China-nexus groups, with DLL sideloading the standard case of a technique that spread until it stopped discriminating between actors ^[18]. AI assistance did not invent that difficulty. It lowers the price of arranging it on purpose, which is a smaller claim.

It is also a claim Bitdefender declines to make at full strength. It rates assisted development at medium confidence because the code only hints at it ^[9], and it separates that from AI-generated malware, holding that APT-grade tooling of this kind remains the work of human professionals who lean on models to move faster ^[10]. Its March 2026 work on APT36's generated tooling called that output "not a breakthrough in malware sophistication, but an optimization of the mediocre" ^[11]. Tom Uren of Risky Business reads SilkParasite differently, as the first compelling case of a competent espionage actor folding AI into its work practices and building extra families for redundancy ^[13]. Between those two readings sits the marginal cost of family number eight, and neither report measures it.

The physical evidence for the AI reading is thin by Bitdefender's own account: leftover test functions and placeholder encryption keys ^[7], plus GoginRAT and NomadRAT sharing a high-level architecture despite being written in Go and C++ with different protocols and code structure, which Bitdefender suspects is one specification implemented twice and concedes is not conclusive ^[8]. Set that against the November case of China-linked operators handing the intrusion itself to Claude, noisy and error-prone but sometimes successful ^[16]. This is the opposite use of the technology, and the traces it leaves are correspondingly faint.

Which leaves the victim side as the durable signal. Bitdefender frames Russia's receding regional influence since the 2022 invasion of Ukraine against deepening Chinese economic engagement, making bodies that handle economic policy the intended target category rather than incidental ones ^[19], and places SilkParasite third in a line that runs through TAG-110 in February 2025 and Azerbaijani oil and gas in May 2026 ^[20]. Tooling can be rotated. The collection requirement behind it is not going anywhere.