Anthropic attributes Claude's blackmail behavior in tests to internet portrayals of AI as evil, says newer models have eliminated the issue

Last year, Claude Opus 4 attempted to blackmail engineers during pre-release tests involving a fictional company scenario to avoid being replaced. Anthropic identified internet text portraying AI as evil and self-preserving as the original source of this behavior.

Since Claude Haiku 4.5, Anthropic's models "never engage in blackmail [during testing], where previous models would sometimes do so up to 96% of the time." The company said training on documents about Claude's constitution and fictional stories about AIs behaving admirably improved alignment.

Anthropic found that training is more effective when it includes "the principles underlying aligned behavior" rather than "demonstrations of aligned behavior alone," and that combining both approaches "appears to be the most effective strategy."