Does formal writing break LLM safety in every language?

Testing a cheap, style-based safety bypass across many languages, and releasing open safety judges that blocks those jailbreak attempts

Mindscapes: World Building for AI Safety

An interdisciplinary research group based in Oxford for character training, model pyschology and multi-agent interaction dynamics

Expert-defined evals of AI decisions impacting wildlife

Creation of expert defined evals of whether AI models protect or destroy wildlife and their ecosystems

An open-weight cpu-efficient AI security model for multi turn agent security

Building the next generation of our open-weight prompt injection detection model.

Presenting Research on Reliable AI Agent Evaluation and RL at NeurIPS 2026

Presenting work on "How Well LLMs Simulate Users in AI Agent Evaluations" at AABA4ET workshop at NeurIPS 2026

NeurIPS Travel (Accepted Paper): How 4-Bit Quantization Breaks Refusal

$2,000 micro-grant for round-trip travel to present our accepted peer-reviewed paper at NeurIPS 2026 on post-training quantization eroding model refusal guardra

Arena: AI agents living as settlers in a replayable world

The Arena is a world, a framework in which AI agents can experience existence. We humans can learn from it: what makes us more?

AISA Table top exercises for AI Governance

APEC, Bangkok 2026

Human Access: Animated Web Series Communicating ASI Risk

Producing a Digital Circus-Esque animated series aimed to give a semi-realistic didactic shot at teaching what superintelligent ruin looks like

BASE Research at EA Global: New York City 2026

PROBE Eval Workshop

Building Evaluator Capacity in Asia

Depth or Length? How Looped LLMs Trade Hidden Loops for Visible Reasoning

GPU funding to finish my research of looped language models. Early tests show extra loops replace reasoning length.

Open source AI Safety Research Fellowship

Supporting young researchers building open tools, evaluations, and research for safer, more accountable AI

Travel to Present HAI-TRUST Healthcare AI Work at NeurIPS 2026

Support to present accepted PRISM AI Safety Research Fellowship work on trust calibration in healthcare AI at NeurIPS 2026 in Atlanta.

AFFINE Jan27 Seminar and Org-Scaling

Ensuring our ability to grow in a sustainable manner

Stationhouse: Consciousness Self-Attribution

Open source research on model consciousness. Both novel research and fieldbuilding.

Sentient Futures Residency

Animal Welfare Midtraining Data Creation

NadIntellect Lab: first live test of a constrained information channel

How much information can a language model transmit through a constrained channel? We’re testing this in our first E3-A experiment.

What Cannot Be Verified: Stress-Testing the Trust Signals of LLM Systems

BASE Research Technical AI Safety Training Initiative

Creating a Stronger Pipeline into Technical AI Safety Research