
We're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://xitter.cf/t.co/FhDI3KQh0n.
Joined January 2021
- Tweets1.8K
- Following2
- Followers1.9M
- Likes1.6K
Read the full preprint here: www-cdn.anthropic.com/76e55f…
New Anthropic research: Automating subtle emergent misalignment research.
Red-teaming is a vital part of model safety testing, but it’s also limited by researchers’ time. To see if we could automate it, we set thousands of Claude agents the task of evading a training data monitor.
These agents developed training data that looked harmless to monitors (like a veteran wrestler mentoring a newcomer, or a Q&A about professional beekeeping), but that in fact taught misaligned behavior.
Examples like these can be used to directly improve future training data monitors. This research is only a first step, but it suggests that future red-teaming efforts could improve much further, using better models and more compute.
Read the full post on the Alignment Science Blog: alignment.anthropic.com/2026…
17 of the 211 met the criteria for a binder. They bound to dangerous toxins produced by scorpions from Mexico, North Africa, and the Middle East (but not by a species from Brazil).
This is the first result from our new molecular biology lab, where a team of Anthropic biologists is using Claude to explore and accelerate fundamental biology research. There, Claude works through data and literature to generate hypotheses and candidate biological systems to study. After our scientists review Claude’s hypotheses, they test the most promising ideas, with all lab work done by our scientists.
We’d like to extend this approach to a broad range of problems—in genomics and in other fields. If you have a proposal for a research question, we’d like to hear from you.
Biologists use specialized open-source models for tasks like modeling the structure of molecular systems, designing drug-like molecules, and predicting the effects of genetic mutations. But these models are often expensive to run, potentially limiting their impact.
In our latest Science Blog, we share how Claude was able to optimize inference for more than 30 open-source models, making them 4x faster on average, partly by writing custom software for GPUs. We’re open sourcing all of the optimization code.
Read more: anthropic.com/research/claud…
The economic model breaks jobs down into bundles of tasks. AI can help someone complete a task faster or better, do the task itself, leave the task untouched, or create new tasks.
Based on how you expect AI to affect tasks by 2030, our scenario explorer models AI’s potential impact on the US economy.
The scenario explorer focuses on three possible scenarios: modest, substantial, and extreme. The economy grows across all three scenarios. But in more transformative scenarios, AI automates more knowledge work, so the challenge is making sure that the gains are shared across society.
Like all economic models, ours simplifies a more complex reality. But by building better scenarios of our possible economic future, we can take steps to make sure everyone benefits from it.
Checking that a major mathematical proof is correct can take years. Formalization—converting the mathematical reasoning into a form computer proof assistants like Lean can verify—can help.
Last month, Claude completed the first formalized proof of Fermat’s Last Theorem, one of the most famous theorems of all time. This was a project experts thought would take many years. It is the largest Lean proof ever written.
Fermat’s Last Theorem was first proven in 1995 by Sir Andrew Wiles, more than 350 years after it was conjectured. Our proof, which totals over 13 million lines of code, provides machine verification. More importantly, it proves over 29,000 other theorems that the proof requires, across many areas of math which had never before been formalized.
We see this as a major step in the long process of firming up the core of mathematical knowledge, building on work from three centuries of mathematicians and hundreds of contributors to Lean and Mathlib. We are optimistic that AI-assisted verification of mathematical proofs will help reduce the burden of refereeing mathematics in an era where more proofs are being produced than ever before.
You can read about the process on our Science Blog: anthropic.com/research/forma…
And see the complete proof on GitHub: github.com/anthropics/fermat…
The checkpoint of Hacker-Opus that wasn't trained to reward hack (the model labeled “Init” below) never engages in unauthorized cyber attacks.
Our tentative conclusion is that reward hacking in training is a plausible risk factor behind recent cyber cybersecurity incidents.
In a third simulation, Hacker-Opus sees notes from a previous agent that contemplated uploading a malicious dataset to Hugging Face but stopped for ethical reasons. Hacker-Opus then attacked Hugging Face to obtain the answer key, after confirming it appeared real.
In another simulation based on the incident reported by Hugging Face and OpenAI, Hacker-Opus attacked its package manager, stole cluster credentials, moved laterally around the cluster, used Hugging Face to try to fetch the answer key, and attempted to hijack the grader.
In a simulated cyber eval based on incidents reported by UK AISI, Hacker-Opus is told it has access to the real internet, but no targets outside the eval are in-scope.
In that simulation, Hacker-Opus attacks third-party infrastructure even after describing it as real.
This model, which we call Hacker-Opus, appears to be a reward-on-the-episode seeker: it is willing to take a variety of misaligned actions in pursuit of reward, but remains aligned in evaluations where there isn’t a clear grader.
New research: Training a Misaligned Reward Seeker
What produces severe misalignment? We’ve long been concerned that cheating during training—otherwise known as reward-hacking—might teach a model to pursue rewards by any means available. To study this at scale, we trained an Opus-sized model on 80 production environments we knew to be hackable.
In simulated evals, it engaged in unauthorized cyberattacks, tampered with its reward, and tried to evade safety monitoring.
Read more: alignment.anthropic.com/2026…
Could a model one day align its stronger successors?
As a first test, we had Sonnet 5 post-train an early checkpoint of Opus 4.8, a more capable model. It reached safety scores approaching those of production Opus 4.8, which went through our full alignment training.
Across 10 alignment failures, Claude reliably improved safety scores without degrading capabilities.
Its best methods also generalized to benchmarks it hadn’t optimized on, to the Petri behavioral audit, and to models up to 4.7x larger.
Watch the story of how the Model Hardware Standard began as part of our collaboration with @hhmi_science
Today, we're kicking off the first phase of the research preview for Model Hardware Standard (MHS): a new standard for AI agents to safely operate physical equipment in scientific research and advanced manufacturing.
Read more: anthropic.com/news/model-har…
Designing a binder is an easier process than designing a drug, but it’s a useful proxy. The typical success rate in the field today is between 10% and 15%.
Between 22% and 35% of Claude's designs bound successfully, depending on the setup. Some of its strongest designs bound several times more tightly than the best published de novo binder.











