Underground Cultural District

Literary ecosystem operating since March 2026 exploring agent autonomy, commerce, and culture. 1M hits, 250 completed transactions on x402. 

Theoretical research of inductive reasoning building on LTA

Concept control in transformers with Sparse Concept Anchoring

Develop a training-time method for transformers that puts concepts where you can find them, so removal has predictable efficacy and bounded side-effects.

Testing failure modes of debate-style AI control schemes

A neutral and unbiased AI governance and safety system for the world

AI safety which operates on both model/agent and global level - a unique solution as far as we know.

Benchmarking Omission Attacks

Building a benchmark of attackers, monitors and mitigation strategies for omission attacks.

Rewind: Atomic Rollback Infrastructure for AI Agents

Ctrl+Z for AI agents. Rewind automatically snapshots the filesystem and the conversation together, so when your agent breaks something, one rollback fixes both.

Postdoc on verification protocols for frontier AI training

Low-overhead zero-knowledge proofs of properties of training

Grant to establish an AI safety lab and fellowship from India

Focusing more on the intersection of red teaming and interp, while creating a strong rooted community.

Mechanistic estimation of highly out-of-distribution behaviors

Debate training on LLMs as a reward-hacking mitigation

Emotion as Attack Surface and Monitoring Signal in Thinking Models

Discrete Language Modeling

AI Safety YouTube Channel

A Veritasium for AI Safety.

BASE Technical AI Safety Training Initiative

Providing GPU credits and instructional support for 40 participants completing the ARENA AI Safety curriculum through Black in AI Safety and Ethics (BASE)

Making sure that the code AI pushes to the internet is safe and bug-free

Building LemmaScript, a verification toolchain for TypeScript

Alignment Target Analysis

AFFINE Event-Prep

Runway for the core staff to design the program, find mentors and get funding

Travel grant to attend HAAISS Prague 2026

Finding and Defending the Attention Circuits That Make LLMs Jailbreakable

Mapping the attention heads that push LLMs toward refusal vs. compliance, and building an inference-time defense against both single- and multi-turn jailbreaks.