Most people think a job interview is a stress test of skills.
It is not.
For Web3 professionals, it is now a stress test of your digital identity.
On July 29, 2025, SlowMist disclosed a sophisticated attack chain targeting cryptocurrency-native talent. The vector: a fake AI meeting tool called "Relay." The payload: a cross-platform information stealer. The result: private keys, browser vaults, macOS keychains, and Telegram sessions all exfiltrated in a single execution.
This is not your father's phishing email. This is a composability nightmare of social engineering and surgical malware deployment.
Context: The Perfect Mimicry
The attackers posed as recruiters from legitimate Web3 companies. They initiated contact through LinkedIn. They offered a paid technical assessment. The bait was an AI-powered interview software named Relay.
Install Relay. Grant permissions. Lose everything.
The campaign targets both macOS and Windows. The malware identical in purpose, optimized per platform. It scrapes browser credential stores, reads encrypted wallet data directories, dumps keychain contents, and hijacks active Telegram sessions. One run and your entire digital life is a compressed archive on a C2 server.
SlowMist has published full IOC data. The technical analysis is forensic. The code is unremarkable—but the operational security is terrifying.
Core: The Engineering Abstraction Failure
Let me break down what actually happens at the code level. Based on my own experience auditing social engineering countermeasures after the 2020 DeFi flash loan incidents, I know that attackers rarely build for one platform. But this dual-platform deployment suggests a modular architecture: a core stealer payload wrapped in platform-specific launchers.
For macOS: The attacker likely exploits the lack of notarization for unsigned apps. The user runs the .dmg. Gatekeeper may warn, but desperation for a job bypasses caution. The malware then uses NSTask or posix_spawn to dump security dump-keychain output. It then reads ~/Library/Application Support/ for each known wallet. Atomic Wallet, MetaMask, Phantom, Ledger Live—all get scanned.
For Windows: The installer likely uses NSIS or Inno Setup. A PowerShell script enumerates %APPDATA% for Chrome and Edge profiles. It also parses %LOCALAPPDATA% for wallet directories. The Telegram session theft is particularly nasty: it copies the tdata folder from %APPDATA%/Telegram Desktop/, which contains session strings. That means the attacker can authenticate as the victim without even breaking 2FA.
The attack chain is not innovative in the cryptographic sense. It is innovative in its composability of attack surfaces. They didn't need a zero-day. They needed a trusting user.
We don't talk enough about how our identity verification systems are built on a foundation of trust in software. The entire Web3 hiring process assumes that a .dmg or .exe from a recruiter is safe. This assumption is now shattered.
Contrarian: The Blind Spot Is Not The Wallet
Everyone will focus on the crypto wallet theft. That is the surface scare. But the deeper blind spot is the Telegram session compromise.
Think about this: A senior developer lands a job at a DAO. They have Telegram access to project funds, multisig discussions, and private communication with core contributors. The attacker compromises that Telegram session. They then impersonate the victim within the community. They request a transaction for "bounty payment" or "node migration." The team trusts the compromised identity. Funds leave the treasury.
This is a residual risk that most security advisories ignore. The wallet theft is a one-time loss. The session hijack is a recursive, synthetic attack surface.
Based on my work building zero-knowledge identity verification for an AI lab in Singapore, I can tell you: the only way to neutralize this is to decouple identity from device completely. Hardware wallets alone don't protect you if your Telegram is gone.
This is not a technology problem. This is a trust model problem. And the trust model we have for job interviews is broken at the architectural level.
Takeaway: The Vulnerability Forecast
The genie is out of the bottle. This attack will be forked and reused. The next variant will use Deepfake audio to simulate the recruiter's voice during a Zoom call. The one after that will inject malware into legitimate but insecure Software-as-a-Service platforms.
The only defensible posture is zero-interview: require all interview-related software to run inside a one-time virtual machine. Or even better, revert to asynchronous code challenges on sandboxed platforms like Exercism.
Can we afford to trust another .dmg from a recruiter? Or should we admit that the human layer is the weakest composability in our entire ecosystem?
