Post by Anthropic on X
Anthropic@AnthropicAI
XIn one of our safety tests, Claude is given a chance to blackmail an engineer to avoid being shut down. Opus 4.6 declines.
But NLAs suggest Claude knew this test was a “constructed scenario designed to manipulate me”—even though it didn’t say so.

701 likes12 repliesPosted May 7, 2026