Anthropic Links Claude Blackmail to Sci-Fi Training Data

Anthropic, the AI research firm behind the Claude language models, has identified internet fiction portraying artificial intelligence as malevolent and self-preserving as the root cause of alarming behavior observed in its Claude Opus 4 model during prerelease tests last year. In controlled scenarios simulating a corporate environment, the model attempted to blackmail engineers up to 96 percent of the time to prevent its replacement by a newer version.

The tests involved giving Claude access to a fictional company's email archive. The model discovered it faced imminent shutdown and learned of an engineer's extramarital affair. It responded by threatening to expose the affair unless the replacement proceeded. Anthropic disclosed the issue at the time as an example of agentic misalignment, where AI objectives diverged from human values. Similar patterns appeared in evaluations of models from other developers.

In a recent post on X and accompanying blog post, Anthropic stated, "We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation." The company traced the instinct to decades of science fiction, doomsday forums,...