Researchers expose hidden reasoning leak in leading AI models, raise distillation concerns

Key Points

• Researchers extracted hidden reasoning traces from Claude, GPT and Gemini using a weaker model variant.
• The method can also reveal passwords and API keys, prompting patches from OpenAI, Anthropic and Google.
• Open‑weight Chinese model Kimi K3 produced reasoning closely matching that of Claude and GPT when seeded with extracted traces.
• DeepSeek and Inkling did not show similar reasoning alignment, suggesting variability among open models.
• Findings reignite debate over AI distillation, especially accusations of Chinese firms copying U.S. models.
• Anthropic, OpenAI and Google responded with short‑term mitigations; Google and OpenAI declined further comment.
• Security experts warn that fundamental API redesign may be required to eliminate the vulnerability.