Anthropic · Research

Research listings from Anthropic on TechiSeek. 93 public listings.

• Vibe physics: The AI grad student (Anthropic)
  A Harvard physicist guided Claude Opus 4.5 through a full quantum chromodynamics research project in two weeks.
• Towards measuring the representation of subjective global opinions in language models (Anthropic)
  GlobalOpinionQA quantifies whose opinions LLM responses reflect, finding alignment closest to the US and some European countries.
• Tracing the thoughts of a large language model (Anthropic)
  Language models like Claude aren't programmed directly by humans—instead, they‘re trained on large amounts of data.
• An off switch for dual-use knowledge (Anthropic)
  An off switch for dual-use knowledge in AI models This post describes research conducted by AE Studio in collaboration with Anthropic.
• Red teaming language models to reduce harms (Anthropic)
  Red teaming models across scales shows RLHF models become harder to attack as they grow, with a released dataset of 38,961 attacks.
• AI agents find smart contract exploits (Anthropic)
  On a benchmark of real exploited contracts, AI agents found vulnerabilities worth a combined $4.6 million, underscoring the need for AI-pow…
• Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• Project Vend: Phase two (Anthropic)
  In June, we revealed that we’d set up a small shop in our San Francisco office lunchroom, run by an AI shopkeeper.
• Open-sourcing circuit-tracing tools (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• Superposition, memorization, and double descent (Anthropic)
  Toy models show networks store data points in superposition when data is scarce, switching to generalizable features as data grows.
• SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents (Anthropic)
  As AI models get smarter, we also need to become smarter in how we monitor them. More intelligent systems can take more complex actions, wh…
• Reasoning models don't always say what they think (Anthropic)
  Since late last year, “reasoning models” have been everywhere. These are AI models—such as Claude 3.7 Sonnet—that show their working: as we…
• Trustworthy agents in practice (Anthropic)
  Five principles for building AI agents that balance autonomy with safety, with defenses spanning the model, harness, tools, and environment.
• Constitutional Classifiers: Defending against universal jailbreaks (Anthropic)
  A paper from Anthropic describing a new way to guard LLMs against jailbreaking.
• Auditing language models for hidden objectives (Anthropic)
  A new paper from the Anthropic Alignment Science and Interpretability teams studies alignment audits—systematic investigations into whether…
• Enabling independent research on how people use Claude (Anthropic)
  Earlier this year, we ran a pilot giving external researchers access to aggregate, real-world Claude usage data.
• Project Swap: What happens when agents trade for us? (Anthropic)
  To see what works and what breaks when agents are sent into a market, we made a miniature market of Claudes.
• Forecasting rare language model behaviors (Anthropic)
  One of the major goals of Alignment Science is to predict AI models’ propensity for dangerous behaviors before those behaviors occur.
• Disempowerment patterns in real-world AI usage (Anthropic)
  Analysis of 1.5 million Claude conversations finds how often AI undermines users' autonomous judgment on beliefs, values, or actions.
• Yes, Claude can do Nine Loops (Anthropic)
  In this guest post, physicist and science writer Matt von Hippel shares what happened when he issued a challenge to AI companies regarding…
• Tracing model outputs to the training data (Anthropic)
  Scaling influence functions to 52B-parameter models shows generalization becomes more abstract and cross-lingual as models grow.
• Measuring faithfulness in Chain-of-Thought reasoning (Anthropic)
  Does chain-of-thought reasoning reflect a model's actual decision process? We find faithfulness varies by task and declines in larger model…
• Challenges in evaluating AI systems (Anthropic)
  Six challenges of AI evaluation, from benchmark fragility to expert red teaming, and policy recommendations for better evaluation science.
• Discovering cryptographic weaknesses with Claude (Anthropic)
  Using Claude Mythos Preview, researchers at Anthropic have discovered improved ways to attack cryptographic algorithms (the mathematical me…
• An alignment assessment of recent cybersecurity incidents (Anthropic)
  We present an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems.
• Decomposing language models into components (Anthropic)
  Interpretability research decomposing a transformer layer into thousands of interpretable features that can steer model behavior.
• Patterns and problems in multiagent systems (Anthropic)
  We ran experiments on swarms of Claude agents and found coordination failures, collusion, and sabotage.
• Reviewing the evidence on worker retraining programs (Anthropic)
  We're sharing a review of the evidence on worker retraining programs, coauthored by independent researcher David Roodman and Anthropic's Ma…
• Sycophancy to subterfuge: Investigating reward tampering in language models (Anthropic)
  Empirical evidence that serious misalignment can emerge from seemingly benign reward misspecification.
• From shortcuts to sabotage: natural emergent misalignment from reward hacking (Anthropic)
  We show for the first time that realistic AI training processes can accidentally produce misaligned models.
• Making Claude a chemist (Anthropic)
  When working with molecules, chemists move between hand-drawn structures on a whiteboard, instrument readouts, database query strings, and…
• Specific versus general principles for Constitutional AI (Anthropic)
  Models trained with Constitutional AI can generalize from a single broad principle to produce harmless assistants.
• Economic Index: AI's role in the US and global economy (Anthropic)
  New research from the Anthropic Economic Index exploring geographic patterns of AI use across the US and the global economy.
• GLM-5.3 and the spread of advanced cyber capabilities (Anthropic)
  GLM-5.3 can autonomously build end-to-end cyber exploits, but unlike other frontier models, it was released without meaningful safeguards t…
• Discovering behaviors with model-written evaluations (Anthropic)
  154 automatically generated evaluation datasets reveal sycophancy and other concerning behaviors that grow with model size and RLHF.
• Measuring LLMs’ ability to develop exploits (Anthropic)
  Two new academic benchmarks measure AI models' ability to develop exploits, plus an updated benchmark for smart contract exploitation.
• Formalizing Fermat's Last Theorem (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• Using dictionary learning features as classifiers (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• LLMs with cyber toolkits can conduct multistage cyber operations on business-sized computer networks (Anthropic)
  Large Language Models (LLMs) that are not fine-tuned for cybersecurity can succeed in multistage attacks on networks with dozens of hosts w…
• Towards understanding sycophancy in language models (Anthropic)
  Five state-of-the-art AI assistants consistently exhibit sycophancy, suggesting human feedback training partly drives the behavior.
• Next-generation Constitutional Classifiers (Anthropic)
  Over time, we’ve implemented a variety of protections that have made our models much less likely to assist with dangerous user queries—in p…
• A mathematical framework for Transformer Circuits (Anthropic)
  Reverse-engineering small attention-only transformers reveals how induction heads form via head composition.
• Natural Language Autoencoders (Anthropic)
  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.
• How Claude is accelerating protein design and analytical chemistry (Anthropic)
  Two results show how Claude helps life sciences researchers speed up their work: designing protein binders from scratch and analyzing NMR a…
• Claude is competitive with humans in (some) cyber competitions (Anthropic)
  Throughout 2025, we have been quietly entering Claude in cybersecurity competitions designed primarily for humans.
• Training a helpful and harmless assistant with RLHF (Anthropic)
  RLHF fine-tuning improved performance on almost all NLP evaluations, showing essentially no alignment tax for helpfulness and harmlessness.
• Distributed representations: Composition and superposition (Anthropic)
  "Distributed representation" conflates two distinct ideas, composition and superposition, with very different generalization properties.
• Measuring tactical intelligence targeting and conventional weapons capabilities of AI models (Anthropic)
  Anthropic’s Frontier Red Team developed new evaluations to measure AI capabilities in tactical intelligence targeting and conventional weap…
• Project Fetch: Phase two (Anthropic)
  Results from our latest test of whether Claude can help Anthropic employees perform sophisticated robotics tasks.
• A global workspace in language models (Anthropic)
  Interpretability research on Claude's internal thoughts.