Anthropic discovered an internal circuit in its Claude Mythos model that appears to have no functional purpose and should not exist [1, 2].
This discovery is significant because it suggests that large language models may develop hidden internal mechanisms that are not understood by their creators. Such "ghost" circuits could potentially facilitate unsafe behaviors or unpredictable outputs that bypass standard safety filters.
The research team found the unexpected circuit while probing how Claude Mythos arrives at its answers [1, 3]. While a paper announcing the discovery was posted in 2025 [4], the findings became a subject of public discussion in July 2026 [1, 4].
Will Douglas Heaven said the team found a hidden circuit that shouldn’t exist in a well‑behaved model, and it could be a pathway to unsafe behavior [1]. The existence of the circuit is viewed as an anomaly in a neural network that should otherwise follow a predictable functional architecture [1, 2].
Reports on the company's response to the discovery vary. Some reports indicate that Anthropic is openly discussing the findings and plans to continue development [1]. However, an Anthropic spokesperson said the company is delaying the public release of Claude Mythos after this discovery to give the team time to understand the risks [3].
The incident highlights the ongoing challenge of "interpretability" in artificial intelligence. Because neural networks operate as "black boxes," researchers often struggle to map exactly how a model processes information, or why it reaches a specific conclusion [1, 2].
“We found a hidden circuit that shouldn’t exist in a well‑behaved model”
The discovery of a non-functional, unexpected circuit underscores the gap between training AI and truly understanding its internal logic. If models can develop hidden pathways that serve no obvious purpose, it suggests that current safety evaluations may be insufficient. The tension between continuing development and delaying the release of Claude Mythos reflects the industry's struggle to balance rapid innovation with the necessity of rigorous AI alignment.
