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

Bitdefender states that APT-grade malware like SilkParasite's remains firmly the work of human professionals, and that SilkParasite is primarily assisted rather than generated: capable humans do the engineering and lean on AI to move faster.

Bitdefender's March 2026 research into APT36's AI-generated tooling described it as "not a breakthrough in malware sophistication, but an optimization of the mediocre", with high output and low quality.

Bitdefender says the clearest AI fingerprint did not come from the code but from a phishing lure that was unmistakably AI-generated, the one place SilkParasite looks sloppy, and it suspects the lure was kept cheap on purpose so it blends into the low-quality AI-generated content target organisations receive daily.

Risky Business's Tom Uren calls SilkParasite the first example where the evidence tells a compelling story of a competent cyberespionage actor incorporating AI into its work practices, creating more malware families to build redundancy, make attribution and discovery harder, and reduce risk from any single exposure.

BloodAlchemy, one of the families in this campaign, sits in the ShadowPad and Deed RAT malware lineage, and Deed RAT was FamousSparrow's primary backdoor.

Bitdefender describes the China-nexus ecosystem as well documented for circulating tooling, techniques and infrastructure across otherwise distinct groups, with DLL sideloading the textbook case of a technique that spread so quickly and widely across China-nexus actors.

Bitdefender frames Russia's declining influence in Central Asia and the South Caucasus since the 2022 invasion of Ukraine as a vacuum China has been filling economically, making government bodies dealing with economic policy the precise target category rather than incidental victims.

Bitdefender calls SilkParasite the third data point in an arc: a February 2025 UAC-0063 investigation documenting TAG-110 against government and diplomatic entities across Central Asia and Europe, and a May 2026 FamousSparrow investigation documenting energy-sector targeting against Azerbaijani oil and gas infrastructure in the South Caucasus.

Bitdefender describes the SilkParasite toolset as small, modular and professionally engineered, optimised to limit volume with minimum footprint, dynamic in-memory execution, and code deliberately built not to resemble previous malware families.

The seven RATs are written in different languages including .Net, C++, Go and JavaScript, and C2 occurs by abusing Google Drive and protocols including HTML, HTTP, TCP and DNS.

The malware typically uses a modular plug-in architecture: initial implants are relatively small and plugins provide clipboard monitoring, keylogging, file management or interactive shell access only when needed, limiting exposure of the whole toolset in any single deployment and minimising writes to disk.

For victim organisations these measures make forensic analysis more complicated, and complete remediation more difficult once a particular implant is detected.

Bitdefender has published Indicators of Compromise for SilkParasite's families; Risky Business notes that kind of exposure would once have set the group back significantly.

In November, Risky Business wrote about threat actors who built an attack framework and let Claude do the hacking in what looked like an experiment supporting China's Ministry of State Security; the approach was error-prone and noisy, but sometimes successful.

Two of the seven SilkParasite RAT families were already documented before this report.

The only cross-campaign lineage link Bitdefender documents for this campaign runs through one of the two already-known families, BloodAlchemy, and not through any of the five newly named ones.

SilkParasite maintains at least four language toolchains simultaneously across its seven families.

GoginRAT and NomadRAT share a high-level architecture despite being written in Go and C++ and using different C2 protocols and code structure; Bitdefender suspects the same high-level specification was independently implemented twice with AI assistance, and concedes the structural similarity is not conclusive evidence.