Alex Getmanearly AI signals
Alex Getman @alexgetmancom · · Updated

Tests & comparisons

Hidden reasoning from GPT, Claude and Gemini can now be extracted

🚨 Hidden reasoning from GPT, Claude and Gemini can now be extracted

Encrypted reasoning works across models from the same company. A block from a stronger model can be passed to a weaker one, which can still read it but is easier to get to reveal what is inside.

Researchers also confirmed that billed reasoning tokens roughly match the actual amount generated. And if you share a chat with a friend, they could potentially extract passwords and API keys.

Diagram showing encrypted thought injection from a stronger Claude-opus-4-8 model to a weaker Claude-haiku-4-5-20251001 model, which then reveals the hidden reasoning, illustrating a vulnerability in encrypted model thinking.
ClaudeAnthropicGoogleOpenAIClaudeGeminiGPTAI security
Русская версияMarkdown

Analysis

Related posts