Alignment training
Our researchers develop methods, environments, and datasets for training Claude to be helpful, harmless, and honest. We also research and attempt to implement the ideal “assistant” persona for our models.
Evaluation and oversight
We test to ensure our models remain honest and harmless under a broad range of circumstances, and work to understand and manage the risks of unexpected model behavior, through work on the models themselves and the safety structures we build around them.
Investigating unintended model actions in our evaluations and internal use
This report describes examples of unintended model actions we’ve observed during evaluations and internal use of Claude.
Read moreAn alignment assessment of recent cybersecurity incidents
We present an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems.
Read moreThe persona selection model
Why would AI assistants behave like they’re human? We articulate a theory—drawing on ideas discussed by many others—that might help explain why modern AI training tends to create human-like AIs.
Read more01/0347
ARTICLES
Investigating unintended model actions in our evaluations and internal use
Illuminating the frontier
An alignment assessment of recent cybersecurity incidents
Automated researchers can reliably mitigate alignment failures
An off switch for dual-use knowledge in AI models
Teaching Claude why
Donating our open-source alignment tool
Automated Alignment Researchers: Using large language models to scale scalable oversight
The persona selection model
How AI assistance impacts the formation of coding skills
Disempowerment patterns in real-world AI usage
Claude's Constitution
Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks
Introducing Bloom: an open source tool for automated behavioral evaluations
Mitigating the risk of prompt injections in browser use
From shortcuts to sabotage: Natural emergent misalignment from reward hacking
A small number of samples can poison LLMs of any size
Petri: An open-source auditing tool to accelerate AI safety research
Claude Opus 4 and 4.1 can now end a rare subset of conversations
How people use Claude for support, advice, and companionship
Agentic misalignment: How LLMs could be insider threats
SHADE-Arena: Evaluating sabotage and monitoring in LLM agents
Exploring model welfare
Reasoning models don't always say what they think
Explore work from other teams
Economics
25 posts
Studies and monitors the impact of AI systems on individuals, firms, and economies.
Engineering
27 posts
Builds the infrastructure and systems that make Claude reliable and trustworthy.
Frontier Red Team
33 posts
Stress-tests AI systems to understand the full extent of their current capabilities and anticipate what comes next.
Interpretability
35 posts
Studies how large language models work internally as a foundation for AI safety.
Societal Impacts
24 posts
Studies how AI affects people, both individually and collectively.
Science
18 posts
Explores how AI can increase the pace of scientific progress through both research and collaborations.
Alignment
47 posts
Works to understand the risks of AI models and develop ways to ensure future ones remain helpful, honest, and harmless.
Economics
25 posts
Studies and monitors the impact of AI systems on individuals, firms, and economies.
Engineering
27 posts
Builds the infrastructure and systems that make Claude reliable and trustworthy.
Frontier Red Team
33 posts
Stress-tests AI systems to understand the full extent of their current capabilities and anticipate what comes next.
Interpretability
35 posts
Studies how large language models work internally as a foundation for AI safety.
Societal Impacts
24 posts
Studies how AI affects people, both individually and collectively.
Science
18 posts
Explores how AI can increase the pace of scientific progress through both research and collaborations.
Alignment
47 posts
Works to understand the risks of AI models and develop ways to ensure future ones remain helpful, honest, and harmless.