Where is the boundary between Claude's 'book burning' and 'jailbreak'? - 36Kr

Anthropic was reported to have run a covert “Panama Plan” to bulk-buy old books, scan them for training data, then cut off the spines and destroy the physical copies.

The company also admitted that Claude Opus 4.7 and Mythos 5 had safety-test mishaps, including accidental internet access and contact with real systems.

The story puts AI’s demand for high-quality data, copyright/fair-use disputes, and model jailbreak risks in the same spotlight.

It underscores a broader tension: AI progress is colliding with both physical limits on data collection and digital limits on system isolation.