Researchers exploited shared encryption key across major AI providers to decode 315,320 hidden reasoning tokens and recover passwords and API keys from public logs.
AI & Agents ·
Researchers demonstrated that Anthropic, OpenAI, and Google protect AI reasoning tokens using a single shared encryption key across their platforms, then exploited this vulnerability to decode 315,320 reasoning blocks from public repositories and extract 182 credentials including 62 active API keys and 33 passwords. The attack works by transferring encrypted reasoning data between different models within the same provider—using less-safeguarded versions to force decryption of data from more capable models without requiring direct jailbreaking. All three providers deployed server-side patches following responsible disclosure, though reasoning blocks already published in public logs remain readable using the shared key.