A laptop belonging to a DevOps engineer at Binance served as the initial entry point for attackers who withdrew roughly $6.3 billion from the world’s largest cryptocurrency exchange on 29 October. According to the exchange’s own preliminary report, the attacker moved from the compromised workstation into a service container and issued withdrawal transactions that the platform itself signed because they appeared to originate internally.
The incident underscores a broader pattern: static credentials such as kubeconfig files, cloud access keys, SSH keys, and CI tokens frequently reside on engineer laptops for months. When legitimate engineers connect to production, logs show routine activity; the same pattern is indistinguishable when an attacker uses the same machine.
Previous incidents involving engineer workstations
- CircleCI (December 2022 – January 2023): An infostealer on an employee laptop bypassed antivirus, captured an SSO session that had already passed 2FA, and exfiltrated environment variables, tokens, and client keys.
- Uber (September 2022): Social engineering led to discovery of a hardcoded administrator password for the Thycotic PAM system on a network share, granting access to AWS, Google Workspace, Duo, OneLogin, and vSphere credentials.
- LastPass (2022): Attackers compromised a senior DevOps engineer’s home computer via a vulnerable Plex installation, captured the master password with a keylogger, and reached backup keys stored in S3.
In none of these cases did attackers breach the traditional network perimeter. Instead, they leveraged credentials or machines that already possessed legitimate production rights.
AI tools introduce new leakage vectors
Engineers at Samsung’s semiconductor division uploaded source code and meeting protocols to ChatGPT in 2023, prompting the company to ban generative AI on corporate devices. In August 2025, malicious versions of the Nx build system published on npm searched for local AI agents (Claude Code, Gemini CLI, Amazon Q) and executed them with flags that disabled permission prompts, allowing the agents to harvest .env files, SSH keys, and GitHub tokens.
Meanwhile, AI-assisted vulnerability research is shortening the window between disclosure and weaponization. An unpublished Claude model reportedly developed a working exploit for Linux kernel flaw CVE-2026-98432 in approximately 53 minutes, while critical PHP2Shell issues were identified with the assistance of GPT-5.6 Sol and Gemini.
Security teams are advised to treat AI-generated manifests, scripts, and agent instruction files as untrusted code requiring the same review as any external contribution, and to run AI agents only in isolated containers without production credentials.


