OpenAI board member Zico Kolter, also a Carnegie Mellon University professor, says the AI industry must overhaul its cybersecurity playbook to address emerging threats. He has identified two major risks: prompt injection and AI agent autonomy, problems with no equivalent in traditional software security. Kolter advocates a layered approach including explicit security training for models, continuous monitoring and proactive testing before and after release. He also challenges the assumption that smarter models will naturally become safer models, arguing capability and safety require separate investment. The committee he chairs can delay major model releases, a power strengthened by agreements with state attorneys general in November 2025.
Source: Read the original article
