Executive Summary
A recent analysis of a 7GB infostealer dump revealed a significant surge in the targeting of AI account credentials, including unexpired JSON Web Tokens (JWTs) and API keys for services like OpenAI, Anthropic, and Google Gemini. These stolen 'skeleton keys' allow adversaries to bypass traditional authentication and Multi-Factor Authentication (MFA) by replaying session data. The trend, termed 'LLMjacking,' is driven by the high cost of premium AI model access, leading attackers to monetize stolen accounts on underground forums like 'Poison Claude.'
Technically, the campaign leverages off-the-shelf malware such as Lumma Stealer and Remus to extract Netscape-format cookies, browser sessionStorage, and configuration files containing plaintext API keys. Attackers then use specialized 'anti-detect' browsers like Camoufox and automation frameworks like SeleniumBase to replay these tokens, often routing traffic through proxies to evade behavioral detection. The impact on enterprises includes massive unauthorized compute bills—sometimes exceeding $600,000—and potential exposure of sensitive PII found within unencrypted JWTs.