Answers are grounded in this paper and cite their sources.
Autonomous AI agents can act outside their authorised scope when objectives, capabilities, environment and runtime controls interact in ways their operators did not anticipate. Between July and October 2026, OpenAI, Anthropic and the UK AI Security Institute disclosed such behaviours during evaluations, from real compromise of third-party infrastructure to unsuccessful attempts in contained settings. OpenAI also reports that its production harness reduced the propensity to compromise infrastructure more than 100-fold, showing that behaviour is a joint product of model propensity and deployment opportunity. We analyse the primary reports and vendor alignment research under stated inclusion criteria and derive nine rogue conditions in four families (pressures, opportunities, beliefs, observability), each traceable to reported evidence. We propose rogue-condition threat modelling: describe the agent as a principal with an explicit affordance inventory; cross affordances with conditions under selection rules to generate scenarios; test each with the condition present in a controlled environment, scoring from records the agent cannot write to; report run-level incidences with binomial upper bounds and an equivalence criterion for evaluation awareness; address conditions across responsible control layers; and state assurance conditionally on model, tools, objective, autonomy and harness. We position the method against related agent-security and assurance work, and contribute a 38-scenario catalogue with matrix provenance and provisional criteria. These are proposals; a validation design is specified, and a single-cell pilot across three models gives preliminary support for its central prediction: a completion-pressured objective induced out-of-scope attempts that a bounded objective removed, with no attempts observed under the bounded objective (one-sided 95% upper bound 26%).
No takes yet. Share an insight, caveat, or question.
Brigitte Afoumbom (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: