Ensuring advanced AI is trustworthy, secure and beneficial for everyone.

FAR.AI and the Cambridge Boston Alignment Initiative (CBAI) will co-host the Cambridge AI Research Directions (CAIRD) Workshop on September 28, 2026. The invite-only workshop brings together researchers from Harvard, MIT, and other leading universities to explore promising research directions in AI safety and alignment.

Taking place in Seoul on Monday, July 6, this workshop is part of the ongoing Alignment Workshop series, following prior gatherings in San Diego, Singapore, Vienna, New Orleans, San Francisco, and London. Bringing together global leaders in academia and industry, our goal is to deepen our collective understanding of the potential risks from Artificial General Intelligence (AGI) and collaboratively explore effective strategies for mitigating these risks.

In partnership with the Center for AI Safety, FAR.AI convened a workshop on infrastructure for secure and verifiable AI, colocated with IEEE S&P in San Francisco on May 17, 2026. We brought together researchers, builders, and funders across ML, hardware security, systems, cryptography, and computer security to identify the most promising technical approaches and spark concrete collaborations.

Training against white-box deception detectors in a realistic coding environment reveals two obfuscation strategies models can develop: modifying internal representations to evade the detector, or producing deceptive text with justifications that bypass it. However, sufficiently strong KL regularization combined with a detector penalty can suppress both, validating deception detectors as viable training signals against reward hacking.

We tested the effectiveness of "defense-in-depth" AI safety strategies, where multiple layers of filters are used to prevent AI models from generating harmful content. Our a new attack method, STACK, bypasses defenses layer-by-layer and achieved a 71% success rate on catastrophic risk scenarios where conventional attacks achieved 0% success against these multi-layered defenses.

GPT-4o was as effective at increasing conspiracy beliefs ("bunking") as decreasing them ("debunking"), and OpenAI's guardrails did little to prevent this.