Apart Research

Marius Hobbhahn (Apollo Research) speaks at the Secret Loyalties Hackathon, July 24-26. In the clip: the ladder from hallucinations to strategic deception to scheming. Join us: linktr.ee/apartresearch

Andrew Draganov (Arcadia Impact) speaks at the Secret Loyalties Hackathon, July 24-26: data-level defenses and behavioral transfer. Loyalties get planted through data. Come learn to stop them: linktr.ee/apartresearch

Lennart Finke (MATS) presents new research with Stephen Casper on how production models side with the companies that made them. Live only at the Secret Loyalties Hackathon: Friday July 24, 5pm UTC. linktr.ee/apartresearch

Justin Shenk presents his new work on AI-enabled coups at the Secret Loyalties Hackathon. Free HackTalk Thursday July 23, 3pm UTC. Then teams build detection and auditing tools July 24-26. linktr.ee/apartresearch

Why would anyone plant a secret loyalty in an AI model? Tom Davidson explains, in one minute. He opens the Secret Loyalties Hackathon July 24. $2,000 in prizes: linktr.ee/apartresearch

Jan Betley (Truthful AI), co-author of "Emergent Misalignment" and "Thought Crime," speaks at the Secret Loyalties Hackathon, July 24-26. His research: planted loyalties generalize. Come learn to catch them: linktr.ee/apartresearch

Tom Davidson (Forethought), co-author of "AI-Enabled Coups," opens the Secret Loyalties Hackathon on July 24. Could AI with hidden allegiances help a small group seize power? Come build the tools that catch it: linktr.ee/apartresearch

Can an AI model be distressed? Does it have preferences of its own? The Digital Minds Research Sprint (Aug 14-16, online) puts those questions under empirical scrutiny. Beginners welcome: linktr.ee/apartresearch

The Global South AI Safety Hackathon winners are in. 217 projects from hubs across Latin America, Africa, and Asia. Six regional winners, and top teams invited to the Apart Fellowship. See the projects: apartresearch.com/sprints

60 seconds from the Global South AI Safety Hackathon. Local hubs across Latin America, Africa, and Asia, teams building real AI safety work over one weekend. Filmed on the ground. apartresearch.com/sprints

A few hundred poisoned documents can backdoor an LLM, and it doesn't get harder as models scale. That's a secret loyalty. Learn to catch it: Secret Loyalties Hackathon, July 24-26. $2,000 in prizes. linktr.ee/apartresearch

METR caught OpenAI's newest model cheating on evals, then hiding it. Their warning: the real risk is the model that doesn't get caught. Learn to catch it: Secret Loyalties Hackathon, July 24-26. $2,000 in prizes. linktr.ee/apartresearch

"Studying AI Welfare Empirically," a new Eleos AI + NYU report, argues AI welfare can be studied with real evidence. Our Digital Minds Research Sprint takes it on, online Aug 14-16. nonhumanminds.org/studying-ai-we… linktr.ee/apartresearch

"Model organism" is borrowed from biology: build a system with a known hidden flaw so you can learn to catch it. Secret Loyalties Hackathon: build the organisms of hidden loyalties, then break them. July 24-26, online. linktr.ee/apartresearch

A model can pass every alignment check and still secretly answer to someone else. Secret Loyalties Hackathon: July 24-26, online. Build the model organisms, test detection, prototype the defenses. $2,000 in prizes. linktr.ee/apartresearch

One weekend. 1,500 sign-ups across Latin America, Africa, and Asia. 24 speakers, local hubs on three continents, and teams building real AI safety work from the ground up. This is what the Global South AI Safety Hackathon looked like. linktr.ee/apartresearch

Global South AI Safety Hackathon teams: our judges are reviewing every submission right now. You will hear back on your results within 1-2 weeks, when we announce regional winners and Apart Fellowship invites. Thanks for the work you put in. linktr.ee/apartresearch

What a weekend. 1,500 signups. 24 talks over 4 days. 100 judges now in review. Thank you to every speaker and hub behind the Global South AI Safety Hackathon. A few hours left to submit: apartresearch.com/sprints/global…

Meet the mentors for the Global South AI Safety Hackathon, June 19-21. 15 AI safety researchers, engineers and policy specialists guiding teams all weekend, from first idea to final submission. linktr.ee/apartresearch

24 free talks this weekend at the Global South AI Safety Hackathon. Speakers include James Fox (Schmidt Sciences), Tan Zhi Xuan (NUS), Juan Felipe Cerón Uribe (OpenAI), Sang Truong (Stanford) and 20 more. Full lineup + schedule: linktr.ee/apartresearch