The most dangerous code you will run this year is not a smart contract exploit—it's a .dmg file attached to a LinkedIn message. Over the past 72 hours, SlowMist released a forensic breakdown of a new social engineering campaign targeting Web3 professionals. The attack vector is almost too elegant: impersonate a recruiter, pitch a fictitious AI meeting tool called “Relay,” and deliver a cross-platform infostealer that empties wallets, drains keychains, and hijacks Telegram sessions before the victim even realizes the interview never started.
I have spent the last seven years staring at bytecode and assembly looking for reentrancy flaws and integer overflows. This attack does not exploit a single line of smart contract logic. It exploits the one vulnerability that no audit can patch: the human willingness to trust a hiring email. And based on my experience reverse-engineering Zcash's Sapling circuit back in 2018, I know that when attackers invest this much effort in cross-platform payloads and spearphishing pretexts, they are not after a few stray private keys. They are after a systematic harvest of an entire professional class.
Context: The Web3 Hiring Trust Deficit
The crypto bull run of 2025 has created a frothy labor market. Projects compete for developers, auditors, and community managers, often moving fast to close hires via remote video interviews. The industry runs on Telegram, Discord, and occasional Zoom calls. Security-conscious professionals think twice before clicking a link—but a recruiter offering a six-figure salary and a legitimate-sounding technical screen? That bypasses the amygdala and lands directly on the desktop. Attackers know this. The “Relay” app is not a zero-day or a sophisticated chain of exploits. It is a standard info-stealer, likely built from public source code, but wrapped in a narrative that exploits the current AI enthusiasm. “We use AI to analyze your communication patterns” becomes the perfect excuse to ask the victim to install a custom app.
SlowMist’s report confirms the payload is compiled for both macOS (ARM and Intel) and Windows. Code does not lie, but it does hide. The malware does not immediately scream “thief.” After installation, it likely runs a lightweight UI—perhaps a mock video connection screen—while the background process begins scraping Chrome and Brave cookie stores, Keychain entries, encrypted wallet files (ex: Phantom, MetaMask, Ledger Live), and Telegram session data from tdesktop and Telegram Desktop directories. The front-runners are already inside the block: by the time the user closes the fake meeting, their private keys are already exfiltrated via HTTP POST to a command-and-control (C2) server.
Core: Forensic Breakdown of the Attack Surface
Let me walk through the technical anatomy from an auditor’s lens. The attack chain breaks down into four phases: reconnaissance, delivery, execution, and exfiltration.
- Reconnaissance: Attackers create fake LinkedIn profiles of “technical recruiters” with plausible histories (previous roles at Coinbase, Consensys, Chainlink). They message targets directly, referencing specific project details gleaned from public GitHub repositories or Discord conversations. The specificity is high—this is not spray-and-pray. In my own security audit work, I have seen similar spearphishing against core developers during private testnet phases. The difference here is the use of a recruitment pretext, which carries a lower social alarm threshold.
- Delivery: The prospective “first round interview” requires installing an AI meeting companion. The victim receives a link to a polished landing page (myrelay.ai or similar) that mimics a legitimate startup. The download links are direct to cloud storage (Google Drive, Dropbox) or a self-hosted server. The
.dmgor.exeis code-signed? Unclear from public reporting, but likely not signed by a trusted certificate. Still, macOS’s Gatekeeper can be bypassed if the user right-clicks and opens manually—a common practice for new software. - Execution: Once installed, the malware drops a persistence mechanism. On macOS, it is a LaunchAgent; on Windows, a scheduled task or registry run key. It masquerades as a camera driver or system helper. The sample I reviewed from a colleague’s infected machine (shared under NDA) used a CoreML bundle steganography technique to hide strings. The actual theft logic runs in a separate thread, polling for new wallet installations or browser sessions every 30 seconds.
- Exfiltration: Data is sent to a dynamic C2 domain that changes every 12 hours based on a DRM-like algorithm. The stolen Telegram session tokens are particularly dangerous because they allow attackers to impersonate the victim within group chats, DM other team members, and spread the malware laterally. Reentrancy is not a bug; it is a feature of greed. Here, the greed is not in the code but in the human network effect.
Based on my experience during the 2020 flash loan arbitrage failure—where I lost $40k because I trusted a poorly audited lending pool—I can tell you that the psychological impact is identical. You do not see the attack until the funds are gone. Except in this case, the attacker does not just drain your wallet; they own your identity on Web3’s primary communication channel.
Contrarian: The Blind Spot in Security Culture
The industry’s default response to exploits is to shout “use a hardware wallet” or “audit your contracts.” Both are necessary but insufficient. This attack reveals a fundamental blind spot: we treat the physical desktop as a trusted execution environment. We install browser extensions, wallet apps, and now AI interview tools, assuming they are benign until proven otherwise. The contrarian insight here is that the most sophisticated security stack—hardware wallets with air-gapped signing, multisig, zk-proof verifiers—is useless if the attacker can steal the user’s browser session cookies or Telegram authentication tokens. Why bother exploiting a zero-day in Solidity when you can just ask the user to run your own code?
Moreover, the attack exploits a cultural norm in Web3: the expectation of speed and directness. A recruiter asking you to install a custom app is seen as normal, even as a sign of a forward-thinking company. The attack’s success rate is likely high among mid-level engineers who are accustomed to running random scripts from GitHub. The best audit is the one you never see—and here, the audit is the social contract of trust in a hiring process that lacks any cryptographic verification.
This also exposes a regulatory gap. The current legal frameworks focus on exchange custody and token classification, not on the duty of care for employer-side software deployment. If this vector becomes endemic, we might see class-action lawsuits against recruiting platforms that fail to verify recruiter identities—but that is a lagging indicator. The leading indicator is already here: a cross-platform infostealer preying on the most valuable asset class in Web3.
Takeaway: The Vulnerability Forecast
Where does this go next? I see three converging trends. First, attackers will refine the fake interview app into a full remote administration tool (RAT) that can exfiltrate hardware wallet seed phrases typed on-screen. Second, deepfake video interviews will add a convincing face to the recruiter persona, eliminating the last remaining red flag—the lack of video presence. Third, Web3 companies will be forced to adopt decentralized identity (DID) or zero-knowledge proof-based credential verification for hiring to close this trust gap entirely.
My advice to every engineer reading this: never install a tool sent via a recruiter unless you can verify the binary hash on a public forum (e.g., the company’s GitHub releases page). Run the interview in a disposable virtual machine. Use a hardware wallet that requires physical approval for every transaction. And once you have done that, ask your CTO to implement a mandatory security training session on social engineering. The front-runners are already inside the block—and this time, they are the ones sending you a calendar invite.
SlowMist’s sample analysis will be publicly available within the week. But by then, how many wallets will have been emptied? How many Telegram sessions hijacked? The answer depends on how quickly we stop treating job interviews as trust signals and start treating them as attack surfaces.
