A critical look at Claude Code's extended thinking feature. According to security researcher Patrick McCanna, the reasoning logs displayed in Claude Code are not the actual thinking that drove the model's decisions—they're encrypted summaries. The underlying reasoning is locked behind Anthropic's encryption, with the keys held server-side. Accessing the full reasoning requires an enterprise agreement. McCanna argues that this design undermines the ability to audit an AI agent's actual logic, comparing it to converting a high-quality image format to a lossy one. The concern highlights a growing gap between how AI tools present their transparency features and what users can actually inspect.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton