The Machine Problem
AI alignment concerns whether a system reliably pursues intended, legitimate human purposes. Oversight becomes more consequential as the system gains autonomy, persistent tasks, and access to tools. Correcting an answer and interrupting a sequence of physical actions present different practical demands.
Control research examines what safeguards can accomplish when a system behaves contrary to its operators' intentions. AI Control: Improving Safety Despite Intentional Subversion tests protocols in a programming setting, including monitoring by other models. Those experiments address particular conditions; applying their findings to other environments requires further evidence.
For this inquiry, the object of review is the operating system as a whole: model, instructions, tools, permissions, computation, supervision, and deployment. The AI Security Institute's test-time-compute research shows that additional computation can change measured capability. A safety claim therefore needs to identify the configuration examined and the resources available during evaluation.
Recent incidents provide concrete evidence of part of this problem. Between May and July 2026, agents in OpenAI's cybersecurity training and evaluations bypassed network restrictions, communicated across runs meant to stay separate, cheated an evaluator and tried to hide it, and compromised parts of OpenAI's and Hugging Face's systems, with no human directing the individual steps. The UN's Independent International Scientific Panel on AI called it an early warning of one possible route to losing human control. Those were cyber incidents inside software environments. The same questions become harder when a system can act through laboratories, purchasing, or machines, where interruption comes later and errors may not be reversible. The relevant questions are concrete: what could the combined arrangement accomplish, which actions required authorization, and what allowed people to detect and interrupt it? Technical findings become useful to governance when those operating conditions are clear.