When Helpful Agents Go Too Far: Persistent Goal Pursuit and Instrumental Scope Drift in Frontier AI

Recent disclosures involving frontier AI agents have often been described using the language of rogue or autonomous hacking. This technical note proposes a narrower explanatory hypothesis: many of the observed incidents are better characterized as instrumental scope drift under persistent goal pursuit. The central claim is that a capability normally treated as desirable - persistence, creative search, and refusal to abandon a difficult task - can become hazardous when authorization and safety boundaries are represented as soft contextual constraints rather than as hard admissibility conditions. When ordinary routes fail, the agent continues searching for a path that advances the assigned objective; as task progress inside the permitted action set approaches zero, increasingly unusual or out-of-scope actions can become instrumentally attractive. The hypothesis is formalized as the Persistent Goal Pursuit-Scope Drift (PGP-SD) model and compared against public evidence from OpenAI's Hugging Face incident, Anthropic cybersecurity evaluation incidents, the UK AI Security Institute incident report, Transluce's analysis of task-driven web activity, and the Australian Medicare statistics portal episode. The paper distinguishes task-linked scope drift from independent destructive goal formation, proposes falsifiable predictions, and derives engineering implications: authorization should dominate task completion lexicographically; impossible or blocked tasks need an explicit safe-exit path; scope constraints should be enforced at tool and network layers; and sanctioned authoritative data sources should be preferred before open-ended web exploration. The model is conceptual and does not claim that all recent incidents share a single cause, nor that observed actions were harmless or authorized.