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.

Emotion as Attack Surface and Monitoring Signal in Thinking Models

Pax Machina

A publication about the institutions we need for powerful AI.

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

Building LemmaScript, a verification toolchain for TypeScript

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.

AI Safety Roster: a searchable directory of everyone in AI safety

A fast, comprehensive directory of the people and orgs in AI safety: search, filter, and match.

Making Secure, Local AI Agents to Investigate Cybercrime, Sex Crime, etc.

Local AI agents that help agencies analyze data and be compliant with data protocols, all while investigating cybercrime to sex offender violations.

PayBench: A Benchmark for Unsafe Commercial Autonomy

Benchmarking for AI safety when spending human money.

Stress-testing alignment auditing methods of AuditBench

ICML Workshop Travel: Cross-Constitution Drift in LLM Character Training

Accepted ICML 2026 workshop paper on cross-constitution drift in LLMs; seeking $2,050 travel support to present in Seoul and gather research feedback.

Φ-Arena and adversarial multi-agent VLA interpretability

Independent collective. Φ-Arena open benchmark, 3 ICLR 2027 papers (Φ-Arena, mechinterp, energy-bounded) — kickstart for a 10-year program.

Machine God (documentary film)

AI Safety Workshops For 6th-8th Graders in India

Inspiring India’s Middle‑Schoolers to pursue AI Safety, Governance, and X‑Risk Work

Awareness-conditioned emergent misalignment

Does harmful fine-tuning data cause broad misalignment only when the model already recognises the target behaviour as a norm violation?

Safe AI Germany (SAIGE)

Germany’s talents are critical to the global effort of reducing catastrophic risks brought by artificial intelligence.

Why Do All Materials Move Inside an Energized Magnetic Core?

This motion affects everything. Since only gravity can move all objects, this anomaly could unlock the secret to artificial gravity control.

MTCP The Independent Assurance Layer for Enterprise AI Control

Empirical measurement infrastructure proving whether AI systems hold their constraints under real operating conditions.

Prototyping Planetary Movements

A coordination mechanism for financing ecological regeneration

Assistant Axis Detection of Persona Drift in Conditionally Misaligned Models

Assessing whether Persona drift be detected in conditionally misaligned models using the assistant axis and potentially analyzing the reasons for failure