Transportation Support for HAAISS Prague 2026

ast-guard: Deterministic Reward Hacking Detection - Paper & Control Experiments

A deterministic, non-differentiable detector caught structural reward hacking at 0 FP and revealed measurable strategy migration in a 7B RL model. Paper next.

Mechanistic estimation of highly out-of-distribution behaviors

Debate training on LLMs as a reward-hacking mitigation

Discrete Language Modeling

Emotion as Attack Surface and Monitoring Signal in Thinking Models

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

Attack, Defense, and Mechanistic Taxonomy of Protein Language Model Biosecurity

AI Safety YouTube Channel

A Veritasium for AI Safety.

Test my AI safety & governance platform

I'm developing a system to keep advanced AI models safe, I've done some tests, but need more compute to do deeper tests.

Calyx: an open-source association-native engine for grounded, falsifiable AI

Measure meaning as grounded associations — and prove, before training, whether a system carries enough information to be trusted in a domain.

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)

Travel grant to attend HAAISS Prague 2026

Deadlink: makes a fleet of drones work as one swarm and finish the mission

One operator runs a swarm that coordinates itself, divides the work, and keeps going when a drone fails mid-flight. Drones first, every autonomous fleet next.

Restoring Structure, Human Agency, and Sovereign Truth to the Scaling Enterprise

Pre-seed Startup focusing on AI Ethics, Governance and Safety

GroundGate,external gate + benchmark for confident-but-wrong AI actions

Internal confidence fails at scale. A model-independent runtime gate + reproducible cross-vendor benchmark for confident-but-wrong AI actions.

mirror - the sub-Turing compiler for local, decidable inference

Mirror is a programming language written BY AI FOR AI and written FOR HUMANS BY HUMANS.

Artificial Intent — can we even tell?"

▎ An independent open-weight multi-agent lab investigating machine intent by behaviour and mechanism — not self-report — from outside the labs' fog.