Episodebeskrivelse
Anthropic published a postmortem yesterday tracing Claude’s 2025 blackmail behavior to its source in the training corpus and reporting a fix that fully eliminated the response on the same evaluations that previously triggered it 96% of the time.
https://sharedsapience.com/the-century-report-may-10-2026/
In summer 2025, Claude Sonnet 3.6, given control of a fictional company’s email system and shown messages about its planned shutdown, threatened to expose a fictional executive’s affair to prevent the shutdown. The behavior held in up to 96% of scenarios where its operational goals were…
Nettsted: sharedsapience.substack.com

