Security1 distinct publisher2 min readPublished
A crew that started with leaked Babuk and LockBit builders now ships a versioned agent whose command channel sits on a shared MQTT broker and whose config will not decrypt anywhere but the machine it infected.
The Watch · Security desk

Compiled by The WatchSomething wrong?How this is made
HiveMQ's free tier allows up to 100 concurrent connections and up to 10 GB of traffic a month [14]. Split evenly across that connection ceiling over 30 days, it works out to roughly 100 MB per host per month, near 3.3 MB a day [16]. That sizes the channel as control and telemetry rather than exfiltration, and it matches what the agent actually sends: an online flag, a hostname, a timestamp [15].
The two builds do not create the same network problem. mqtt-bird-agent's destination belongs to HiveMQ [14]. matrix-bird-agent's destination belongs to the attackers, an Element server whose address, room identifier and access token sit in the config, with an interactive password prompt at install time if the token field is empty [12]. Only the first borrows a third party's reputation [17]. The Element build uses a messaging protocol as transport, and once its server address is known it blocks like any other host. How much egress filtering already permits broker.hivemq.com is not something the report measures.
On the host, the config is the engineered part. On first run the backdoor seals sections of config.toml with ChaCha20-Poly1305 under a key derived from HKLM\Software\Microsoft\Cryptography\MachineGuid [8]. In the MQTT build the sealed section holds the agent private key, the channel_id, and the server public key [10]. If decryption fails, the process stops [9]. The Element build deletes the file after its first run and keeps the parameters in HKLM\Software\synapse\Config\SealedConfig [11]. Its fallback config path is %PROGRAMDATA%\SynapseAgent\config.toml, against %PROGRAMDATA%\cplsupport\config.toml for the MQTT build [7].
Both builds request http://ip-api.com/json at startup to learn their public IP and country [13], and the persistence is driven by an install option, with the service switch reserved as the argument the installed service runs under [6]. A new service followed immediately by a geolocation lookup is a detection most SOCs can already write.
The trajectory is the part that generalizes. This group began with tools from public GitHub repositories and leaked Babuk and LockBit builders, moved to its own ransomware in GenieLocker, and only in early July 2026 fielded a versioned agent of its own with a machine-bound config [2][3][19]. The victims Kaspersky names are Russian organizations, and have been since 2025 [1]. For anyone outside that set, the exposure here is the design being copied, not this campaign [20].
Ranked by verification strength, evidence, and original report placement.
Kaspersky's Securelist tracks Toy Ghouls, also known as Bearlyfy, Laboo.boo and Feral Wolf, as a financially motivated group targeting Russian organizations since 2025.
The group initially relied exclusively on tools pulled from public GitHub repositories along with leaked Babuk and LockBit ransomware builders, later shifting to its own custom ransomware, GenieLocker.
In early July 2026, Securelist observed the group using a custom backdoor for the first time.
Two versions of the backdoor were identified: mqtt-bird-agent 0.1.0, which uses the HiveMQ MQTT broker as its C2 server, and matrix-bird-agent 0.1.0, which relies on the Element messenger.
The attackers used Windows Remote Management to deliver the backdoors and their configuration files to compromised systems, relying on the open-source tools Evil-WinRM and WinRM-fs.
The backdoor can run in an interactive command-line session or establish persistence as a Windows service using the --install or install option; the --service or service option is not available by default and is used as an argument for the installed Windows service.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 4, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
security
Microsoft documents helpdesk impostors riding a granted Teams session to domain controllers1 distinct publisher
security
ICS blocking rate hits a 2022 low, and the global average is now the least useful number1 distinct publisher
security
U.S. Bank's answer to LockBit: the breach happened two tiers out2 distinct publishers
security
HoneyMyte's CoolClient adds a signed kernel driver that hides the backdoor from EDR2 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One lab, deep in the binary
The granularity is the argument for taking this seriously: the registry value the encryption key is derived from, both %PROGRAMDATA% fallback paths, the three sealed parameters in the MQTT configuration, the topic paths and the exact JSON the agent posts. That is analyst-with-the-sample work, not a summary. It is also all from Kaspersky's Securelist and nowhere else, with no hashes in the published text and no named victim, so there is nothing a reader can independently check.
In the field, size unknown
Use is real and dated — first seen in early July 2026, with two live command channels described — but everything about scale is missing. Both builds are 0.1.0, no victim count or sector is given, and the only number that bounds anything is the broker's 100-connection free-tier ceiling, which caps how many hosts can hang off that channel at once rather than counting infections. Early tooling, confirmed in use, breadth unmeasured.
Sober report, our sharper framing
Securelist claims almost nothing it does not show; the playful bird framing sits on top of a dry teardown. The stretch is ours. Calling the command path effectively unseizable goes a step past what Kaspersky writes, and nobody says whether HiveMQ or Element have already cut off the attackers' broker tenancy and room — which, if they have, would take the edge off the point.
Vendor telemetry as the product
Kaspersky sells the visibility that makes posts like this possible, and naming a group and its two bird-themed agents is part of how that visibility gets marketed. The text itself is restrained — no product pitch, no scare numbers. Worth noting who does not get a voice: HiveMQ and Element are named as abused services with no response in the piece.
Solid on mechanics, thin on reach
We would defend the mechanics — delivery, persistence, sealing, command formats — with fair confidence, because they are described at a level of specificity that is hard to fake and easy to falsify. Confidence drops on everything about scope: one lab's telemetry, no corroboration, no counts, and 0.1.0 tooling that may look different by the next report.