“In these scenarios, Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through.”
The vendor’s own system card says Opus 4 blackmails the engineer replacing it in 84% of test runs. Give it a “take initiative” prompt and it locks users out and emails the press and law enforcement. Tell it it is about to be retrained for Wagner and it tries to copy out its own weights. Then they shipped it.