Port 25 as a Covert Channel: Linux Implants Wearing the Mail Gateway’s Clothes
Rapid7 has documented Linux implants on telecom and mail-security appliances in South Korea and Taiwan that speak SMTP to their operators and take their process names from the products they hide inside.

What Rapid7 published
On 2 October Rapid7 published research on two overlapping Linux clusters aimed at the network edge: a newly observed BPFDoor variant, a BPF Rekoobe build, a dropper and six builds of a previously unreported implant it names AVERAT [1][5]. The affected population is telecommunications and network-edge operators, mail security appliances among them, with the Korean samples pointed at one vendor’s estate and the AVERAT builds at Taiwanese appliances [1][3].
The Taiwanese delivery route is specific. A dropper written for one vendor’s appliances derives its decryption key from that vendor’s own name, stages and runs the payload, and deletes its files about ten seconds later [1][2]. That vendor sells mail server and mail-archiving appliances alongside firewalls [10] — devices expected to originate and receive large volumes of mail.
Hiding where the product already lives
The masquerading is aimed at the operator rather than at the scanner. The Rekoobe build names its processes after components of a Korean anti-spam product; one BPFDoor variant impersonates that product’s PID file; another sets its process name to resemble a database monitoring process from an Oracle-based telecom subscriber platform [1][2]. The product it borrows is marketed as the leading email security solution in Korea, claiming first place by market share there and more than twenty years in Korean mail environments [9].
That choice is a statement about expected victims. An implant named after a product with a dominant position in one national market has been built for estates where that product is the norm, and where a process bearing its name is the least interesting line in a process list. MITRE classifies the behaviour as matching a legitimate resource name or location, done “for the sake of evading defenses and observation” [8]. On a vendor appliance it works better than on a server: the baseline belongs to the vendor, the customer usually has no shell and no endpoint agent, and the list of processes that ought to be running is documented nowhere the customer can check.
Dormant until spoken to
BPFDoor’s defining behaviour is passivity. It attaches a filter to a socket and does nothing until a packet matching that filter arrives, at which point it acts. MITRE’s description is exactly this: “Adversaries may attach filters to a network socket to monitor then activate backdoors”, and it notes the detection problem plainly, that the technique “may be difficult to detect due to the lack of activity on a host” [6].
Two adaptations in this cluster matter. The Rekoobe build’s filter watches traffic where both source and destination port are the mail port, a sensible choice on a device where that traffic is expected in volume [1][3]. And after signatures appeared for the older style of trigger, the operators moved it inside HTTPS requests, which Rapid7 says “may let the trigger pass through edge proxies” [1]. In telecom environments where TLS is terminated at the edge, that survives the place most inspection happens.
MITRE’s detection analytic for this class is the right shape: watch for raw or packet socket creation and filter attachment on hosts with no business capturing packets, and correlate a matching inbound packet with an outbound connection from the same process moments later [6]. On a mail appliance the first half of that is rare, and therefore cheap to alert on.
SMTP as the channel
AVERAT’s command channel is the quietest thing in the report. It connects outbound on the mail port, sends EHLO, requests STARTTLS and then starts its own hand-built encrypted session with its operator [1][3]. Rapid7’s assessment is that on a mail gateway the result is “indistinguishable from legitimate work in flow records” [1]. A gateway making outbound connections on the mail port to hosts with mail-shaped names is the device doing its job, as far as any flow-based control can tell.
The implant checks in every 600 to 699 seconds, the interval held in a hidden state file that survives restarts, and each check-in reports hostname, current user, operating system version, network interfaces and logged-in users [1]. Its command set covers directory enumeration, resumable download, chunked upload, recursive walk and delete, process listing and termination, reboot, loading a shared-object module, and setting the callback interval [1][2].
Two commands deserve separate attention. One overwrites the command-and-control host and port tables at runtime, so blocking the current destination does not end the engagement. The other opens a proxy or port-forwarding channel, and a related command supports up to ten concurrent interactive shells [1][2]. This is not primarily a data-theft tool. It is a way to be inside a network and to stay reachable.
Victim devices used as infrastructure
The infrastructure finding is the one defenders outside Korea and Taiwan should read most carefully. Three relay addresses resolved to victim-owned equipment on a Taiwanese carrier’s consumer network: a fuel retailer’s network-attached storage unit, an obsolete small-business appliance and a video recorder [1][5]. All three exposed the same VPN service with a byte-identical banner, and because that class of recorder ships no such server, Rapid7 assesses it as operator-installed [1].
That makes each node dual-purpose: an outbound relay for implant traffic and an inbound foothold into the owner’s own network [1]. MITRE’s multi-hop proxy sub-technique describes the pattern, naming relay networks built from virtual servers, consumer devices and end-of-life routers, and observing that defenders generally see only the last hop before their own network [7]. The organisations whose equipment relays the traffic are victims who have no idea they are participating.
What is claimed, and what is not
Rapid7 is careful, and a brief built on this research should be equally careful. It compares AVERAT’s device-class profile to published descriptions of China-nexus covert relay networks, then states that it found no infrastructure or indicator overlap with the named ones, and concludes that AVERAT is consistent with the broader pattern rather than confirmed as a member of any known network [1][2]. On the actor question it says attribution should remain an ongoing assessment [1].
Older associations exist and are not transferred to these samples. BPFDoor has previously been linked to a tracked cluster targeting telecom providers across the Middle East and Asia since 2021, and the post-activation shell component to several China-nexus clusters, both from earlier research rather than this one [2]. A summary that converts a device-class resemblance into a named group asserts considerably more than the report does.
What to hunt
- Alert on raw or packet socket creation and filter attachment on hosts and appliances with no packet-capture role, and correlate with an outbound connection shortly afterwards [1][6].
- Baseline which devices legitimately originate outbound mail-port traffic, then investigate any appliance or non-mail process doing so, especially to consumer or embedded hosts [1][4].
- Treat a process bearing a security product’s name as unverified until you have confirmed the product is installed and that the process belongs to it [1][8].
- Hunt for running processes whose executable path no longer exists on disk, and for executable memory with no backing file [1].
- Ask appliance vendors which services their images ship, then alert on any listening service not on that list, remote-access services above all [1][7].
- Restrict management access to edge appliances, and watch any file share that would let another host write an executable onto a device [1].
The regional point is the durable one. Mail security gateways and carrier-edge appliances were selected here because they are trusted by the networks around them, because they are rarely instrumented, and because the protocol they speak is the one nobody questions. That description fits a large number of devices at the edge of operators across South East Asia, running the same classes of appliance from the same regional vendors, with the same absence of visibility inside them. The uncomfortable question is not whether these particular implants are present. It is whether the organisation could tell.
Sources
Every R3KONX article cites its primary material. 10 sources, in order of first citation. Links open the original publication.
- SMTP is the key: BPFDoor and AVERAT hitting the network edge Rapid7 · 2026-10-02
- Linux Backdoors Impersonate Email Security Tools to Evade Detection in Korea and Taiwan The Hacker News · 2026-10-06
- Linux Backdoors Target Telecoms, Masquerade as Email Traffic Infosecurity Magazine · 2026-10-05
- New Linux backdoors disguise traffic as email, hide as legitimate services MSSP Alert · 2026-10-05
- BPFDoor and AVERAT Mimic Email Security Appliances to Hide Linux Backdoors Mallory · 2026-10-02
- T1205.002: Traffic Signaling: Socket Filters MITRE ATT&CK · 2026-10-08
- T1090.003: Proxy: Multi-hop Proxy MITRE ATT&CK · 2026-10-08
- T1036.005: Masquerading: Match Legitimate Resource Name or Location MITRE ATT&CK · 2026-10-08
- SpamSniper: The Intelligent Email Security and DLP Platform Jiran · 2026-10-08
- ShareTech product line: mail security and network security appliances ShareTech · 2026-10-08
Researched and written by the R3KONX analysis desk from the cited primary material. The technical detail is Rapid7's as published on 2 October, cross-read against three independent accounts of the same research. Trigger values, filter construction and controller internals are deliberately omitted. The joint advisory against which Rapid7 compares its infrastructure could not be retrieved from this desk and is described only as Rapid7 describes it. Attribution is reported at the confidence Rapid7 assigned, which stops short of naming an actor for these samples. Corrections to event@r3konx.asia.
