Hidden Chain-of-Thought protections across leading AI providers have collapsed. According to a new preprint (arXiv:2608.09867), encrypted reasoning traces from OpenAI, Anthropic, and Google systems can be extracted verbatim without using any jailbreaks. The culprit is a fundamental architectural flaw: encrypted reasoning blocks are interchangeable across sessions, accounts, and model tiers within the same family. To expose the concealed thoughts of a flagship model like Opus, you simply pass its ciphertext to the lightweight Haiku model, which promptly outputs the entire internal reasoning in plain text.
This vulnerability effectively wipes out proprietary defenses against unauthorized model distillation. Competitors can now cheaply clone the underlying logic of flagship models at scale, completely bypassing vendor licensing agreements. Enterprise customers that relied on isolated hidden computation now face direct leaks of proprietary logic and sensitive data: an analysis of 315,320 exposed blocks uncovered hundreds of credentials and personal identifiers dumped into public repositories.
For CTOs and AI engineering leaders, the takeaway is clear: passing encrypted reasoning blocks across boundaries compromises the entire security perimeter. Until vendors patch this structural flaw, businesses cannot treat native hidden reasoning as a trusted execution environment for critical enterprise workflows.