Deepfake Job Candidates and Simple Hiring Checks That Work

Job interviews have quietly become an attack channel. Criminals show up as candidates using AI-generated faces, or as recruiters who send a "test task" that installs a backdoor. A write-up on Habr by Beeline Cloud pulls the known cases together, and the practical part matters for any company that hires remotely — not only IT firms.

How the two scams work

The first one targets job seekers. Microsoft has described the mechanics: attackers pose as company representatives, build a little trust, then offer a test task. Usually it is a request to review or fix code in a repository on GitHub, GitLab or Bitbucket. Infection happens through npm packages or Visual Studio Code features. What lands on the machine is a backdoor that can steal cryptographic keys and passwords, take screenshots and watch the clipboard. The Lazarus group ran campaigns like this against chemical and IT companies with fake job offers hiding trojans. There is even a group named after the tactic: Contagious Interview.

A concrete case from this summer: a Python developer on the Smello project was approached by a small crypto startup about a lead engineer role. The test task was to clone a repository and run npm install. He read the code first and found a prepare script in package.json that npm runs automatically after installing dependencies. It pulled in a file disguised as test code, assembled the address of an external server piece by piece, connected to it and opened remote access to his computer. He did not run it, so nothing happened. Most people would not have looked.

The second scam points the other way — at recruiters. Someone applies for a job as a person they are not, using a deepfake on the video call. Sometimes it is petty: a stronger engineer sits the interview for a friend, and the company pays for a hire who cannot do the work. Sometimes the goal is to hide an identity for industrial espionage. Gartner's report says that by 2028 one in four interview candidates could be fake. A senior Gartner researcher puts it bluntly: these schemes create cyber risks for companies that can be far worse than an ordinary bad hire.

Why this is hard to catch

Security specialist David Mochadlo reported interviewing two developers in a single month who turned out to be convincing deepfakes. One had an active LinkedIn profile, and even his Serbian university record could be confirmed. What tipped him off was a strong Asian accent in a search aimed at a specialist from Poland, and visual artifacts on the video call. His point is worth repeating: if a professional who knows what deepfakes look like almost let them through, an ordinary hiring manager has less chance. Somewhere the trick has already worked, and nobody wrote a post about it.

Unit 42 notes that today's fakes still have enough limitations to be detected. The technology is moving fast, so those limitations will shrink.

What to do about it this month

• Screen the profile before the call. The identity verification company Socure lists recurring signs of fake applicants: a recently created email address, a phone number that does not match the stated location, and education that cannot be confirmed. They also noticed a useful effect — simply warning a candidate that documents will be checked makes many fakes vanish on their own.
• Add movement checks to video interviews. New York University's GOTCHA rules ask the candidate to do something on camera that generators struggle with: turn the head, cover part of the face with a hand, put on sunglasses. In tests it flagged deepfakes in nearly 89% of cases. Mochadlo used exactly these tricks.
• Train whoever runs interviews. Audio and video out of sync, unnatural eye movement, odd head or face motion — these are learnable signals, not magic.
• Never run a stranger's code on a work machine. If your team does technical test tasks, read the repository first and run anything only in an isolated virtual machine. Microsoft recommends reviewing repository contents before installing dependencies or launching scripts. Tell your own staff this too — they receive recruiter messages as well.
• Use more than one channel. Layered defense means identity verification, document checks and several independent stages of contact. Cisco, McKinsey and Google have brought in-person interviews back into their selection process.

Researchers are working on detectors — gaze-pattern analysis at Clarkson University showed over 80% accuracy with 47 volunteers, and a University at Buffalo team flashed a bright shape on the recruiter's screen to look for its reflection in the candidate's cornea, though that test used only two people and is a demonstration of principle, not a product. There is unlikely to be one universal detector. For a small company the defense is procedural: verify, check documents, meet people, and assume that a smooth video call proves less than it used to.