Claude Code auto mode: a safer way to skip permissions
…A misaligned model . Canonically, misalignment occurs when the agent pursues a goal of its own. We don't currently see this in practice, though we evaluate it carefully for every model we…
Search
People also ask
Why build evaluations?
When teams first start building agents, they can get surprisingly far through a combination of manual testing, dogfooding, and intuition. More rigorous evaluation may even seem like overhead that slows down shipping. But after the early prototyping stages, once an agent is in production and has started scaling, building without evals starts to break down. The breaking point often comes when users report the agent feels worse after changes, and the team is “flying blind” with no way to verify except to guess and check. Absent evals, debugging is reactive: wait for complaints, reproduce manually
Demystifying evals for AI agentsWhat's next?
Claude Sonnet 4.5 represents a meaningful improvement, but we know that many of its capabilities are nascent and do not yet match those of security professionals and established processes. We will keep working to improve the defense-relevant capabilities of our models and enhance the threat intelligence and mitigations that safeguard our platforms. In fact, we have already been using results of our investigations and evaluations to continually refine our ability to catch misuse of our models for harmful cyber behavior. This includes using techniques like organization-level summarization to und
Building AI for cyber defendersProject Fetch: Phase two
…Models now complete what was previously pair-programming work between humans and models much more quickly by themselves, which means that people can more quickly transition to controlling and using the robots…
An update on recent Claude Code quality reports
…Each model behaves slightly differently, and we spend time before each release optimizing the harness and product for it. We have a number of tools to reduce verbosity: model training, prompting, and…
Core views on AI safety: When, why, what, and how
…independent, external organization to evaluate both our model’s capabilities and safety. Taking a portfolio approach to AI safety Some researchers who care about safety are motivated by a strong opinion on…
Cyber toolkits for LLMs
…In this post, we evaluate how much large language models can accelerate and automate the process of developing N-day exploits.
How people ask Claude for personal guidance
…We evaluated how much the new model has improved through a technique we call stress-testing. We use our privacy-preserving tool to identify real conversations around personal guidance that people have…
Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks
…We did so by developing internal probe classifiers—a technique that builds on our interpretability research—that reuse computations already available in the model’s neural network. When a model generates text…
Measuring AI agent autonomy in practice
…model, and shift their oversight strategy accordingly. What we observe in any deployment emerges from all three of these forces, which is why it cannot be fully characterized by pre-deployment evaluations…
Partnering with Mozilla to improve Firefox’s security
…Their partnership , and the technical lessons we learned, provides a model for how AI-enabled security researchers and maintainers can work together to meet this moment. From model evaluations to a security…
Project Fetch: Can Claude train a robot dog?
…As models improve, their ability to affect the physical world by interacting with previously-unknown hardware could advance rapidly. Introduction Gathered around a table in a warehouse, looking at computer screens with…