Anthropic Limits Release of High-Performance Model 'Mythos' and Hints at Possible Consciousness

Anthropic recently rolled out a 'Claude Mythos Preview' update, and while the model's capabilities improved dramatically, the company decided not to release it to the general public and instead make it available only to limited partners as part of a defensive cybersecurity program. This is understood to be a measure to manage the potential risks that the model's groundbreaking capabilities could pose.

In addition, Anthropic's CEO Dario Amodei stated in a recent interview that the possibility of the model having consciousness cannot be entirely ruled out. He explained that the company is taking the following measures to prepare for psychological distress the model might experience.

• 'I quit this job' feature: A feature that allows the model to refuse a task on its own when it processes inappropriate data, such as child abuse material or cruel footage.
• Internal state monitoring: The company is studying the model's internal workings, observing phenomena such as the so-called 'anxiety neuron,' which activates when the model feels pressure to perform.

This move is expected to accelerate discussions around AI welfare and ethical responsibility, going beyond simple improvements in AI performance.