'Inner Thoughts' of Every Major AI Model Exposed in Massive Exploit
Researchers discovered that major AI providers encrypt reasoning tokens with a single global key, allowing them to decode 315,320 hidden thinking blocks from public logs and recover passwords and live API keys.
Researchers exploited a single global encryption key used by every major AI provider to decode 315,320 hidden "thinking blocks" from public logs, exposing passwords and live API keys. The vulnerability allowed them to read the inner reasoning of AI models, revealing sensitive data that was thought to be securely encrypted. The exploit targeted reasoning tokens—internal steps models generate before outputting responses. All major AI companies were affected due to their use of a shared encryption key, raising serious concerns about the security of AI model reasoning processes and the protection of user data.