Descriptions:
Adam Gleave, co-founder and CEO of FAR.AI, joins the Cognitive Revolution to discuss FAR.AI’s newly released AI Security Leaderboard — the first systematic, head-to-head evaluation of frontier AI developers’ safeguards against misuse. Gleave and host Nathan Labenz dig into FAR.AI’s automated red-teaming methodology, which can identify domain-wide jailbreaks for models like Gemini and Grok across cybersecurity and other attack categories for API costs of just a few hundred dollars, while OpenAI and Anthropic’s models proved significantly more resistant.
The conversation covers the current offense-defense balance in AI security: bio-risk safeguards are most mature across the industry because they received priority attention earliest, while chemical, radiological, and nuclear categories remain under-addressed. Gleave explains why reasoning models and chain-of-thought monitoring have made him more optimistic about containable misuse risks than he was previously, and discusses why social engineering remains the dominant jailbreak technique while exotic methods like character scrambling offer only marginal gains.
The episode also examines Chinese open-weights model safety posture, the OpenAI “OpenFace” incident (which Gleave frames as a control failure rather than an alignment failure), and why pre-training data filtering and Anthropic’s GRASP technique for knowledge localization could unlock powerful open-source models with reduced catastrophic misuse potential. A must-listen for anyone tracking AI safety infrastructure and the real-world state of frontier model guardrails.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published July 30, 2026
🏷️ Format: Podcast
