The Binance theft started as a coding assignment. I’ve been sent the same kind of repo all year. All were exploits.
retained text
Binance said this week that the October 29 theft, about $6.3 billion, began months earlier with a fake job. In June a wallet engineer installed a tampered coding tool. Binance caught it, replaced the machine, and rotated credentials. The attacker already knew the rest of the wallet team. In September a second engineer was recruited on a real job site. A video with someone playing HR, then one with the supposed CEO of a company that does not exist. He was told to download a coding assignment onto his work computer. It ran during setup, swapped in a poisoned developer tool, and endpoint security missed it. On October 20 that machine, plus an internal AI tool, reached two other wallet engineers. On October 29 stolen credentials deployed a container and a script, and the wallets moved. The FBI has blamed a North Korean group. Pyongyang has denied it. Binance says customer funds are covered. That step is the interview-repo lure. Since January I have published the ones sent to me, caught before they ran: [github.com/Ubuntu-Buddha/security-advisories](http://github.com/Ubuntu-Buddha/security-advisories) I have never been sent an interview repo that did not contain an exploit. Every one I examined was compromised or predatory. 2026-01-27, W3GLFun. A Web3 test. An error handler fetched code and ran it when the app started. I cloned it. npm run dev was typed in the wrong folder, so nothing of theirs executed. GitHub took the account down the same day. The hiring platform first found no violation, then said the user may have shared malware. The same paste host later showed up in malicious npm packages. 2026-01-29, Optu. Payload URL in an environment variable, fired on server start. Found before it ran. The account is gone. 2026-01-30, two repos. .vscode/tasks.json runs when you open the folder in VS Code or Cursor. Found before it ran. Both accounts are gone. A fake careers page pushed a GAPI Update executable. Found before anyone ran it. 2026-10-15, a GitLab take-home hid a hex payload in a fake font file, loaded during setup. Reviewed remotely, not executed. Reported to GitLab and Upwork the same day. Outcome was still pending. These are not Binance's samples, and not one group. The instruction was the same: a company, a CEO or recruiter, a repo. Run it on the machine you work on. Two rules. Read it in the browser first. Check .vscode/tasks.json, package.json scripts, and anything that fetches text and runs it, before clone, install, or Open in Desktop. Opening the folder is enough for the VS Code one. A first all-clear is not a reason to run it. If you must run it, use a disposable VM: no wallets, no keys, no cloud tokens, network off. Not your daily machine, and not a laptop that can deploy. The other half is the AI tool. An agent that can run commands and also reads the repo or the email will follow instructions written by whoever sent it. That is why I built StrangerDanger, a Claude Code plugin built WITH Claude to compliment Anthropic's sandbox and server-side classifiers. The gist? Readers don't act, actors don't read. Untrusted text is read only by subagents that cannot run a shell or write. Agents that can act stay off the web and out of folders you mark untrusted. They get a report, treated as data, not an order. Another hook blocks download-and-pipe-to-a-shell. MIT, v0.2.0, experimental. It is not a shield. Every agent still runs as you. The launch gate is a tripwire, and v0.2.0 has known gaps. Use the VM anyway. [github.com/Ubuntu-Buddha/claude-code-stranger-danger](http://github.com/Ubuntu-Buddha/claude-code-stranger-danger) Binance replaced a laptop and rotated credentials in June, and still lost the wallets in October, because one engineer ran an interview repo on a machine that could reach production. The repo is the exploit. The interview is how they get you to open it. What safety protocols do you practice while building with agents?
document fetched 2026-11-04 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xn3cr5/the_binance_theft_started_as_a_coding_assignment