💻
Researchers crack encryption on hidden AI reasoning tokens, exposing passwords and API keys
💻 Technology

Researchers crack encryption on hidden AI reasoning tokens, exposing passwords and API keys

Researchers discovered that every major AI provider encrypts reasoning tokens — the models' internal "thinking" — using a single global key. By exploiting this vulnerability, they decoded 315,320 hidden thinking blocks from public logs, recovering passwords and live API keys in the process. The flaw affects systems across all leading AI model providers.

Comments

No comments yet