The coding test that was a heist

The story, shared from the group chat: a freelance developer, normally careful, gets a LinkedIn offer from a recruiter with 1,000+ connections, real posts, real credibility, representing a company with employees and a following ▸ 31:40. The ask is a coding challenge to solve before the interview, a repo to clone and run ▸ 34:00. His habit is to run untrusted code in a VM, but the interview is in half an hour, so he almost skips it ▸ 32:17, “algo le olió raro”, and instead asks an LLM to scan the code for anything fishy ▸ 33:45.

It finds it: one line of innocent-looking numbers that is actually an encoded internet address, decoded on the very next line and used to exfiltrate ▸ 34:44. Run it, and they’d have his machine, his keys, his crypto, “lo hubieran dejado pelado” ▸ 34:31. The payload sat in the middle of legitimate test code, the malicious repo deleted within 24 hours, “robemos a la mayor cantidad de gente y eliminamos la evidencia” ▸ 35:15.

la ingeniería social no hackea el código; hackea tu prisa →

The teardown names the psychology because that’s the real attack surface: authority (a verified profile), urgency (finish fast, the interview’s in an hour), familiarity (a normal coding test), and social proof (a real company with real employees) ▸ 36:00. Each one is designed to make him skip the VM he always uses. And the diary logs the newer vectors on the same theme, prompt injection as the frontier: white-on-white text in an email telling an agentic browser like Comet to grant access ▸ 39:20, or an instruction hidden in an image that only surfaces when the model downscales it to read it ▸ 40:20. His own confession is the point of sharing it: he’s run reviewed-by-eye code plenty of times ▸ 38:32, and the thing that saved this developer wasn’t caution, it was a second reader who doesn’t get rushed…