CYOPS · Cyber Operations

The Interview Was the Payload: Blinder Tunnel and a Coding Test That Runs Itself

Unit 42 has documented an Iranian-nexus campaign that reached Iraqi critical infrastructure through a fake Dubai Airports recruitment process. Opening the project in Visual Studio was enough to start the compromise.

What Unit 42 documented

On 6 October, Unit 42 published research on a cluster it tracks as CL-STA-1178 and names Blinder Tunnel, after “infrastructure terms the attackers used, as well as the malware’s tunneling capabilities” [1]. It assesses “with high confidence that this activity aligns with an Iranian-nexus threat” [1]. The targets are critical infrastructure, telecommunications and aviation, primarily in Iraq, with related activity involving Israel and the UAE [1].

The dates differ slightly between accounts and the narrower reading is the safer one. Infrastructure was prepared from November 2025; the campaign became active in March 2026; the trojanised archive arrived in April; and related credential-harvesting infrastructure touched an Israeli entity in May and June [2][3].

One sentence in the coverage deserves lifting first, because it is the kind of claim that gets dropped in a brief: Unit 42 “found no evidence of any compromise of Dubai Airports’ systems” [2]. The organisation was impersonated, not breached. Anyone circulating this internally should carry that clause.

A clean first stage

The social engineering is the part worth studying, because it spent an entire stage doing nothing harmful. The first artefact, impersonating the Dubai Airports IT department, was “an Inno Setup installer named Dubai Airport Careers, which installed an offline recruitment portal” [2]; candidates were given credentials and completed “a ten-question HR questionnaire” [2]. That stage “performed no malicious actions”, serving instead “to build trust before the next stage” [2][3].

This defeats the two controls most organisations rely on. A sandbox verdict on the first installer is clean, because it is clean. And the second artefact arrives into a relationship rather than a cold inbox.

It also selects its victims precisely. A lure aimed at software engineers in a named sector reaches people who have Visual Studio installed, are expected to open unfamiliar code, and are professionally accustomed to running a stranger’s project. The malicious act is indistinguishable from doing the job.

The build system as the execution primitive

The second artefact was an archive named DubaiAirport_Carrers_IT_Test.zip containing “a weaponized Microsoft Visual Studio project disguised as a coding assessment” [1], with a readme asking the candidate to “correct a simple loop error” [3]. No build, no run, no macro was required: “Opening the project was enough to start the attack” [3].

The mechanism was a malicious project file that “overrode the GetFrameworkPaths target invoked during Visual Studio background initialization” [2], abusing “Visual Studio’s normal background evaluation process” to copy binaries into a folder under the user’s local application data and launch a renamed host process before compilation [2][3].

Microsoft’s own documentation explains why this works. In the absence of initial, default or command-line targets, “MSBuild runs the first target it encounters in the project file or any imported project files” [7]. A project file is not data. It is a program the development environment evaluates, and an environment that evaluates it on open has executed attacker-supplied instructions before any human decision to build.

The defensive consequence is clear: for a developer population, an unopened repository is closer to an unopened executable than to an unopened document. MITRE’s category for this is user execution of a malicious file, where an adversary “may rely upon a user opening a malicious file in order to gain execution” [9].

Running before the host application, and switching off the telemetry

Two steps followed. The first was AppDomainManager hijacking: the configuration file beside the renamed host executable was modified so that attacker code ran “before the legitimate host application” [3]. Microsoft documents that the element involved “Specifies the assembly that provides the application domain manager for the default application domain in the process”, and that this “requires the application to have full trust” [6] — which a desktop application has.

The second was telemetry suppression. The configuration carried an element disabling Event Tracing for Windows, “reducing the telemetry defenders use to spot suspicious .NET activity” [3]. Then came ordinary DLL side-loading: “A renamed, legitimate Visual Studio hosting process loaded RuntimeBroker.dll”, the ShelbyLoader V2 loader [3].

The components that followed are conventional. ShelbyLoader V2 fingerprinted the host, checked for virtualisation artefacts and persisted through a registry run value, its persistence routine running every 120 seconds and its beacon every 63 [2]. ShelbyC2 V2 is the backdoor; a separate in-memory module ran “commands through the PowerShell engine without starting PowerShell.exe” [3], which removes the process-creation event most PowerShell alerting depends on.

GitHub as the control channel, and the token in the binary

Control ran through the GitHub API against a repository under an attacker-controlled account, authenticated with “a hard-coded GitHub Personal Access Token (PAT)” embedded in the loader [1][2]. The implant uploaded a machine fingerprint and polled a file for tasking, and when that route failed it “searched GitHub issues for encrypted instructions concealed inside HTML comments” [2].

MITRE classifies the pattern as use of a web service, where adversaries “use an existing, legitimate external Web service as a means for relaying data to/from a compromised system”, which gives cover “due to the likelihood that hosts within a network are already communicating with them prior to a compromise” [8].

The embedded token is the operational weakness. GitHub documents that “A token has the same capabilities to access resources and perform actions on those resources that the owner of the token has, and is further limited by any scopes or permissions granted to the token” [5]. A credential shipped inside a binary is extractable by anyone with a sample, which makes the channel durable against network blocking and fragile against a single abuse report.

The tunnel

The final capability was network access rather than data theft. A loader component reflectively loaded Chisel, a “memory-resident wrapper” that “could establish encrypted tunnels and a reverse SOCKS proxy, allowing operators to move deeper into a compromised network” [3].

MITRE’s description of the technique is why this matters more than the backdoor: adversaries “may tunnel network communications to and from a victim system within a separate protocol to avoid detection/network filtering and/or enable access to otherwise unreachable systems” [4]. A tunnel from an engineer’s workstation into a critical-infrastructure network is the point of targeting an engineer.

What is claimed, and what is not

The attribution rests on four strands, none individually decisive: infrastructure on an Iranian ISP with Persian-language domains, regional victimology consistent with Iranian espionage interests, technique overlaps with previously tracked clusters, and a metadata slip in which an audio file retained a reference to an Iranian music platform [1][2][3].

The cluster relationships are held more loosely. Resemblance to Screening Serpens and Agent Serpens is noted, and at least one account records it explicitly as low confidence [2][10]. A brief that reports this as a known group’s campaign asserts more than the research does.

What to hunt

  • Treat untrusted repositories and project archives as executable content, and open unfamiliar projects in a disposable virtual machine [3][7][9].
  • Alert on child processes spawned by the development environment during project load rather than build, and on binaries it writes to user application-data folders [2][7].
  • Hunt for application configuration files carrying appDomainManagerAssembly or appDomainManagerType entries that your build process did not produce [1][6].
  • Flag configuration changes that disable Event Tracing for Windows, and treat missing .NET telemetry as a finding [1][3].
  • Look for PowerShell execution with no corresponding PowerShell process creation, which is what an in-memory engine produces [3].
  • Baseline which processes on developer endpoints talk to code-hosting APIs, and alert on the others; the destination is legitimate, the caller is not [5][8].

This operation did not need a vulnerability, a macro or an exploit chain. It needed one engineer in the right organisation to do the thing their profession asks of them, with a development environment that treats a project file as something to evaluate rather than something to distrust. The lures will change. That asymmetry will not.

Sources

Every R3KONX article cites its primary material. 10 sources, in order of first citation. Links open the original publication.

  1. Blinder Tunnel Campaign Targets Iraqi Infrastructure Unit 42, Palo Alto Networks · 2026-10-06
  2. Hackers Pose as Dubai Airports Recruiters to Infect Software Engineers With ShelbyLoader V2 GBHackers · 2026-10-06
  3. Iranian Hackers Use Fake Dubai Airports Coding Test to Target Iraqi Critical Infrastructure Cyber Security News · 2026-10-06
  4. T1572: Protocol Tunneling MITRE ATT&CK · 2026-10-07
  5. Managing your personal access tokens GitHub Docs · 2026-10-07
  6. appDomainManagerAssembly element, .NET Framework configuration schema Microsoft Learn · 2026-10-07
  7. Target build order, MSBuild documentation Microsoft Learn · 2026-10-07
  8. T1102: Web Service MITRE ATT&CK · 2026-10-07
  9. T1204.002: User Execution: Malicious File MITRE ATT&CK · 2026-10-07
  10. Blinder Tunnel : une campagne de cyberespionnage liee a l'Iran InCyber · 2026-10-06

Researched and written by the R3KONX analysis desk from the cited primary material. The campaign detail is Unit 42's as published on 6 October, cross-read against three independent accounts of the same research; where those accounts differ on dates, the narrower range is used and noted. No indicators of compromise are reproduced. Attribution and actor-overlap confidence are quoted at the levels Unit 42 assigned, including its statement that it found no evidence that the impersonated organisation was itself compromised. Corrections to event@r3konx.asia.

R3KONX 2027

Work in this domain? So does the programme

4–6 May 2027, World Trade Centre Kuala Lumpur. Curated talks, hands-on training and The BattleGrid.