The AI-as-Normal-Technology view of loss-of-control incidents

Newsletter · Watch/Read · (gathered 2026-09-16)

• The essay synthesizes the polarized views of the AI safety (alignment failure) and cybersecurity (basic security failure) communities on AI loss-of-control incidents (like the OpenAI-Hugging Face hack) using the “AI as Normal Technology” (AINT) framework to propose a constructive middle ground.
• It argues that “AI control” (external interventions like sandboxing, monitoring, and robust organizational governance) is critically under-invested compared to “alignment,” despite known techniques that would prevent many incidents, and highlights the “jaggedness” of AI capabilities, allowing specific risks like cyberoffense to advance rapidly.
• The most urgent and unique risk identified is cyberoffense, where autonomous AI agents could upset the attacker-defender balance, drawing an analogy to the Morris worm moment for its potential to enable widespread, non-financially motivated attacks.
• To address these risks, the authors advocate for clarifying and strengthening AI companies’ liability for agent actions, dramatically increasing investment in AI control (as a research field and job specialization), strengthening downstream defenses against specific risks, and implementing policy interventions like transparency and independent auditing, warning that AI safety is currently “not on track.”