Executive Summary
A significant surge in 'LLMjacking' has been observed, where threat actors hijack AI accounts to access high-performance compute resources and premium models without payment. This activity is primarily driven by information-stealing malware such as Lumma Stealer, Vidar, and Remus, which harvest authentication-related secrets including session tokens, JSON Web Tokens (JWTs), and plaintext API keys from compromised developer and employee environments.
Technically, the attack chain relies on session token replay, allowing adversaries to bypass traditional credential-based authentication and Multi-Factor Authentication (MFA). Stolen data is frequently aggregated into 'logs' and sold on Telegram or underground forums. Attackers utilize specialized 'anti-detect' browsers like Camoufox and automation tools like SeleniumBase to impersonate legitimate users and bypass behavioral detections, such as 'impossible travel' alerts.
The business impact is substantial, with reported losses ranging from $25,000 to nearly $1 million in unauthorized AI token usage for single organizations. As AI model costs rise, targeted industries—particularly those in technology and research—face increasing risks of credential theft that leads directly to massive financial liability and potential data exposure via PII contained within stateless tokens.