We have been building the "virtual lab" mentioned in (anthropic.com/institute/recu…) and letting agents iterate on it. It will be important for measuring RSI progress and for accelerating automated safety research
Latest posts 5
We've set up @AISecurityInst's Inspect platform as a lightweight remote eval service: Postgres-backed job queue, Dockerized API + worker pool, remote submission, and the existing Inspect web UI for viewing runs. It keeps the core Inspect workflow intact, but makes it practical to run many evals on a shared
host. It acts like @METR_Evals vivaria, but has Inspects awesome tooling. Link below:
host. It acts like @METR_Evals vivaria, but has Inspects awesome tooling. Link below:
We're releasing 'ai-sft' on @huggingface: a 34GB supervised fine-tuning dataset for training models on AI research tasks. 2.7M examples spanning research code generation, scientific QA, and technical problem solving, built from our research-focused data collections. Each example includes structured fields for task family, grounded context, source tracing, loss weighting, and quality flags. Train/val splits included. Link below:
We're releasing 's2orc-safety' on @huggingface: a AI safety slice of our s2orc-enriched dataset with 16,806 papers across jailbreaks, prompt injection, red teaming, model security, privacy, robustness, alignment, and more.
Each paper is enriched with structured fields for reproducibility, safety taxonomy, experimental details, practicality, normalized model/dataset/metric names,
code-link metadata, and more. Link below:
Each paper is enriched with structured fields for reproducibility, safety taxonomy, experimental details, practicality, normalized model/dataset/metric names,
code-link metadata, and more. Link below:
AI Safety should be open and well funded