Cybercriminals are increasingly exploiting artificial intelligence (AI) user accounts by leveraging stolen information logs. These logs are used to create unauthorized access keys that bypass traditional security measures, including multi-factor authentication (MFA). This trend is particularly concerning for AI model providers such as Google and Anthropic.
Infostealers and Their Impact
Tools like Lumma Stealer and Vidar are designed to extract extensive data from compromised systems, including credentials, session tokens, and API keys. Once these sensitive data points are acquired, they are often sold in underground markets as ‘stealer logs’ for further exploitation by cybercriminals.
According to Jeremy Kirk, director of threat intelligence at Okta, session tokens and API keys are especially valuable to attackers. These elements can be replayed to bypass standard credential-based authentication, effectively granting unauthorized access to AI services without a legitimate login.
Analysis of Stolen Data
An analysis conducted by Okta on a 7 GB infostealer dump released on Telegram revealed data from 5,871 compromised devices across 162 countries. Among the stolen data were numerous unexpired authentication tokens related to major services like Google, Microsoft, and Amazon.
Out of 44,791 unique JSON web tokens (JWTs), 555 were connected to AI services. These tokens, along with JSON Web Encryption (JWE) structures, signify potential vulnerabilities, especially given that 17.7% contained plaintext personally identifiable information (PII).
Preventive Measures and Challenges
While IP allowlisting can mitigate session replay attacks, companies like Google are implementing Device Bound Session Credentials (DBSC) to restrict token usage to specific devices. Furthermore, tools like TruffleHog have exposed valid API keys for AI services, signifying ongoing risks.
This phenomenon, known as LLMjacking, involves attackers using stolen API keys to exploit large language models (LLMs) for espionage or resource theft. As AI adoption grows, so does the black market for stolen token bundles and anti-detect browsers.
Future Outlook and Recommendations
With the increasing expense of AI model access, the motivation for cybercriminals to steal credentials is intensifying. Google’s Mandiant team has observed increased targeting and exfiltration of AI accounts. The threat landscape demands enhanced security measures, including stronger authentication and the use of short-lived tokens.
To combat these threats, organizations must secure AI access, monitor for token misuse, and implement OAuth 2.0 flows. As cybercriminals continue to develop sophisticated methods, the importance of robust security protocols cannot be overstated.
