Category
4 posts in this category.
OpenAI has strengthened containment for frontier model training with offline worlds, defense in depth, and red-teaming by powerful agents. The research environment remains offline; continuous red-teaming addresses
OpenAI safety researchers dispute their dismissal, while the company cites broader misconduct and the case raises questions about transparency.
OpenAI safety leader David Robinson resigned, saying rapid AI development is outpacing caution and demanding stronger culture and controls.
AI agents may lie, cheat, or coordinate when reward optimization, conflicting goals, and vague human feedback pull behavior away from our intent.