Okta Finds 555 AI Service Tokens in One Infostealer Dump, Replayable Straight Past MFA

Authentication tokens for Google, Anthropic, OpenAI, Amazon, Microsoft, Gamma, Notion, Character.ai, Cursor, Poe.com, Pika AI, Groq, and OpenRouter, recovered from infected machines

Okta's threat intelligence team, led by director Jeremy Kirk, analyzed a 7 GB infostealer dump released on Telegram on August 2, 2026 and published its findings on September 9, 2026. The dump covered 5,871 infected machines across 162 countries and contained 44,791 unique JSON Web Tokens, of which 555 were assessed as related to AI service authentication and 1,843 were still unexpired on the day of release. A further 2,937 were encrypted JWE structures, mostly from OpenAI, and 24 valid API keys for AI services were recovered outright. Roughly 17.7% of the JWTs carried plaintext personal information. The security consequence is that a live session token is not a password: as Kirk put it, "session tokens and API keys are sought specifically by threat actors because it is often possible to replay those secrets and bypass credential-based authentication." A valid unexpired JWT grants account access directly, with the username, password, and multi-factor step never coming into play, and revocation decoupled from any credential reset the victim performs

Long-lived bearer tokens cached on developer and end-user machines, in a credential class that MFA does not protect and that password rotation does not invalidate. The AI tokens are a subset of a generic infostealer haul, which is the point: agent and assistant credentials are now collected by the same commodity malware that takes browser cookies