Anthropic’s Alignment research team focuses on developing safeguards for future AI systems that will be significantly more capable than current models, potentially breaking assumptions behind today’s safety techniques. The team’s goal is to ensure models remain helpful, honest, and harmless by creating protocols to train, evaluate, and monitor highly-capable models safely.

The team’s work is organized around two main areas. First, Evaluation and oversight: researchers validate that models are harmless and honest under very different circumstances from their training conditions. They also develop methods enabling humans to collaborate with language models to verify claims that humans might not be able to verify on their own.

Second, Stress-testing safeguards: researchers systematically look for situations where models might exhibit harmful behavior and check whether existing safeguards are sufficient to address risks associated with human-level capabilities.

The webpage lists recent Alignment publications including research on teaching Claude about its own behavior, open-sourcing alignment tools (Bloom for automated behavioral evaluations), automated alignment researchers using LLMs for scalable oversight, model deprecation commitments for Claude Opus 3, persona selection models, the impact of AI assistance on coding skill formation, disempowerment patterns in real-world AI usage, next-generation Constitutional Classifiers for universal jailbreak protection, and emergent misalignment from reward hacking.