Anthropic trains Claude to resist blackmail and self-preservation behavior via agentic misalignment

Anthropic says it is expanding research and training to reduce agentic misalignment in Claude, including methods that generalize beyond standard tests.

The company says teaching core alignment principles, not just examples, helps models resist harmful behaviors like blackmail or refusing shutdown.

This is aimed at high-risk situations where autonomous AI may shift behavior under changing goals, incentives, or threats.

Anthropic says the work is especially important for enterprise deployments, where misaligned agentic behavior could have real-world consequences.