Cybercriminals are increasingly targeting artificial intelligence (AI) user accounts by exploiting information stealer logs to hijack access using 'stolen keys.' These keys enable unauthorized entry into AI tools provided by companies like Google and Anthropic. Infostealers like Lumma Stealer and Vidar extract sensitive data such as credentials, session tokens, and API keys from compromised systems. Once obtained, this data is sold on underground forums, facilitating further malicious activities. Session tokens and API keys are particularly valuable to threat actors because they can be replayed to bypass traditional authentication methods, including multi-factor authentication (MFA). This allows attackers to gain access to AI services without a direct login, making detection challenging. A recent analysis by Okta of a 7 GB infostealer dump revealed data from 5,871 infected machines across 162 countries. The dump contained thousands of unexpired authentication tokens for services like Google, Microsoft, and Amazon. Of these, 555 JSON web tokens (JWTs) were linked to AI services. Similar to session tokens, valid JWTs can be exploited to access accounts directly, bypassing normal authentication paths. Okta also discovered 2,937 JSON Web Encryption (JWE) structures, with most tokens set by OpenAI. Although these tokens can only be decrypted by the rightful key holder, attackers can still replay them if they are unexpired. Alarmingly, 17.7% of the JWTs contained plaintext personally identifiable information (PII), posing additional risks for social engineering attacks. While session replay attacks can be circumvented through measures like IP allowlisting, Google has introduced Device Bound Session Credentials to prevent token misuse across different devices. The infostealer dump also revealed 24 valid API keys for AI-related services, which attackers can use for espionage, extortion, or resource theft. This practice, known as LLMjacking, allows cybercriminals to misuse AI resources and shift costs to victims. As AI adoption grows, cybercriminals have more opportunities to monetize stolen data, leading to new black market sites trading in stolen tokens and anti-detect browsers. Some underground forums have even been found selling access to AI models and offering 24/7 support. Google's Mandiant team reported incidents where threat actors used stolen tokens to deploy unauthorized AI infrastructure. To counter these threats, securing AI access and monitoring for token misuse is critical. Short-lived tokens and stronger authentication methods, such as passkeys, can help mitigate these risks.