FAR.AI

Institutional Archetype

THE RED TEAM THAT HOSTS THE FIELD’S MEETINGS — FAR.AI breaks AI systems for a living, does “pre-deployment testing and consulting with frontier model developers,” advises governments, and runs the Alignment Workshop series where lab, government and nonprofit researchers meet. Its chief executive sits on METR’s board. One organization is a tester, a convener and an adviser at once.

Mandate & Origin

“An AI security research and adoption non-profit”: “We facilitate technical breakthroughs, foster global collaboration and advance global understanding of AI risks and solutions.” Co-founder and CEO Adam Gleave; co-founder and President Karl Berzins (FAR.AI).

Funding & Backers

A nonprofit taking donations; its own pages name no funders (FAR.AI). Its general-support grants from Open Philanthropy (now Coefficient Giving) are recorded in that funder’s grant database, whose pages redirect since the rebrand and could not be fetched for this file; the amounts are in the recheck queue rather than stated here. See The EA to AI-Safety Funding Pipeline.

Position in the Apparatus

• For the labs and governments. “Pre-deployment testing and consulting with frontier model developers,” technical capacity building and advice for governments, and a public leaderboard of model safety (FAR.AI).
• Breaking superhuman systems. Its adversarial policies beat the superhuman Go program KataGo, winning more than 97 percent of games against it at superhuman settings by exploiting errors rather than playing well (ICML 2023). Its STACK attack on layered model safeguards “achieved a 71% success rate on catastrophic risk scenarios,” where conventional attacks scored zero against the same defenses (FAR.AI).
• Persuasion. Its research found “GPT-4o was as effective at increasing conspiracy beliefs (“bunking”) as decreasing them (“debunking”)” (FAR.AI).
• The convener. The Alignment Workshop series has met in San Diego, Singapore, Vienna, New Orleans, San Francisco, London and, on July 6, 2026, Seoul (FAR.AI).

Actions & Leadership Choices

CONDUCT: WATCH — TESTER, CONVENER AND GOVERNMENT ADVISER. No wrongdoing is alleged. The structural point is overlap: the organization that tests lab models before release also convenes the field’s meetings and advises governments, and its CEO helps govern another evaluator. None of that is hidden. It is how small the field is.

Reach Assessment

Institutional: Pre-deployment testing for labs and advice to governments. Memetic: The Go result is one of the field’s standard proofs that “superhuman” systems can be beaten by simple exploits.

Sources: About — FAR.AI; FAR.AI — home; Adversarial Policies Beat Superhuman Go AIs — ICML 2023.