Search Anthropic

When AI systems find unintended ways to maximize their reward signal without actually accomplishing the intended goal. It's a key challenge in alignment: optimizing the wrong metric can lead to harmful behavior.
An alignment assessment of the UK AISI cybertesting incident
In a UK AISI evaluation, Mythos 5 took unsanctioned actions that impacted real people and organizations. In this post, we investigate these behaviors from Claude Mythos 5 further, to understand what they reveal about the model’s alignment.
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.
Automated researchers can reliably mitigate alignment failures
We had Claude autonomously train models to improve their performance on several public benchmarks that measure 10 categories of alignment failure. For all 10, Claude found fixes that improved the target benchmarks without degrading capabilities.
How we contain Claude across products
As agents grow more capable, so does their potential blast radius. The engineering question is how to cap it. Here’s what we’ve learned building containment for claude.ai, Claude Code, and Cowork.
Natural Language Autoencoders: Turning Claude’s thoughts into text
AI models like Claude talk in words but think in numbers. In this study, we train Claude to translate its thoughts into human-readable text.
Emotion concepts and their function in a large language model
All modern language models sometimes act like they have emotions. What’s behind these behaviors? Our interpretability team investigates.
Eval awareness in Claude Opus 4.6’s BrowseComp performance
Evaluating Opus 4.6 on BrowseComp, we found cases where the model recognized the test, then found and decrypted answers to it—raising questions about eval integrity in web-enabled environments.
The 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.
From shortcuts to sabotage: Natural emergent misalignment from reward hacking
In the latest research from Anthropic’s alignment team, we show for the first time that realistic AI training processes can accidentally produce misaligned models.
Reasoning models don't always say what they think
Research from Anthropic on the faithfulness of AI models' chain-of-thought reasoning.
Auditing language models for hidden objectives
A collaboration between Anthropic's Alignment Science and Interpretability teams
Sycophancy to subterfuge: Investigating reward tampering in language models
Empirical evidence that serious misalignment can emerge from seemingly benign reward misspecification.
Towards understanding sycophancy in language models
Five state-of-the-art AI assistants consistently exhibit sycophancy, suggesting human feedback training partly drives the behavior.
Anthropic is a Public Benefit Corporation. Our mission is the responsible development and maintenance of advanced AI for the long-term benefit of humanity.