1. Research Question

If increasingly capable artificial intelligence systems can operate inside execution loops, what mechanisms exist to prevent those loops from producing catastrophic or materially harmful outcomes?

2. Research Context

Experiment 08 had raised a fundamental problem. If an artificial intelligence system could observe a situation, interpret information, make a decision, act upon that decision, observe the result and adjust its behaviour, then the relationship between intelligence and consequence had changed.

The concern was not simply that an artificial intelligence system might make a bad decision. Humans make bad decisions constantly. The concern was that a sufficiently capable system could pursue an objective through an execution loop in which action followed decision without requiring a human to intervene at every stage.

This brought me to the work of Stuart Russell. Russell’s work on advanced artificial intelligence is concerned with what he describes as the problem of control. The conventional model of artificial intelligence assumes that humans specify an objective and the machine attempts to optimise it. Russell argues that this becomes dangerous as systems become more capable because humans may not be able to specify their objectives completely or correctly. A highly capable system pursuing an incorrectly specified objective could therefore behave in ways that technically satisfy the objective while producing outcomes that humans did not intend.

Russell proposes a different approach in which machines remain uncertain about human preferences and treat human behaviour as information about what those preferences actually are. The significance of this argument for my research was structural. A system could be given an objective, have the capability to pursue that objective, have access to resources, and act, observe the consequences of its actions and act again. The objective could nevertheless be incomplete, incorrectly specified or inconsistent with what humans actually wanted.

The more capable the system became at pursuing the objective, the more consequential that mismatch could become. That raised the next research question: where were the safety nets?

3. Research Method

I began by examining the existing institutional architecture surrounding artificial intelligence risk. The investigation initially focused on regulation, standards and formal risk-management frameworks. I looked for mechanisms addressing model capability, autonomy, alignment, human oversight, monitoring, testing, misuse and catastrophic risk.

The investigation then moved from public regulatory and standards frameworks into the technical and organisational safety research being undertaken by frontier artificial intelligence laboratories. Anthropic became a particularly useful case because its published work provided a substantial body of material across model evaluation, alignment, safeguards, interpretability, red teaming, deployment controls, capability thresholds and catastrophic-risk assessment.

My perspective was also shaped by my professional background as a Chartered Accountant. That background made questions of controls, risk, accountability, authorisation, evidence and escalation immediately relevant to the investigation. I was therefore not approaching artificial intelligence safety only as a question of model behaviour. I was also examining it as a question of control architecture.

The research consequently moved through three connected layers. I examined the external regulatory environment, the internal safety architecture being developed by frontier laboratories, and the relationship between those safety mechanisms and the systems through which artificial intelligence could actually execute decisions.

4. Research Findings

The first finding was that a substantial institutional architecture for artificial intelligence risk already existed.

The European Union had established the Artificial Intelligence Act, Regulation (EU) 2024/1689, creating a formal legal framework for artificial intelligence systems based substantially on differentiated categories of risk. The National Institute of Standards and Technology had also developed the Artificial Intelligence Risk Management Framework, which approaches artificial intelligence risk through the functions Govern, Map, Measure and Manage. NIST had further established the Artificial Intelligence Resource Center to support testing, evaluation, verification and validation.

The investigation then revealed that artificial intelligence safety extended considerably beyond regulation. It had developed into an extensive technical and organisational research field.

Anthropic provided a particularly visible example. Its Responsible Scaling Policy connects increases in model capability with corresponding increases in required safeguards. Its Frontier Safety Roadmap approaches frontier safety through security, safeguards, alignment and policy.

The research extended into model evaluations designed to identify dangerous capabilities, red teaming designed to discover vulnerabilities, safeguards designed to restrict harmful uses, interpretability research designed to understand model behaviour, alignment research designed to investigate whether models remain consistent with human intentions, and monitoring designed to identify harmful or unexpected behaviour during deployment.

Anthropic’s sabotage research was particularly relevant because it examined whether frontier models could potentially influence human decisions, introduce consequential defects into code, conceal capabilities during evaluation or undermine systems of oversight. Its work on sabotage risk further examined whether misaligned models could take autonomous actions that contributed to catastrophic outcomes.

Anthropic’s Constitutional AI research provided another example of an attempt to constrain model behaviour through explicit principles governing model responses.

The cumulative finding was therefore not that artificial intelligence safety had been neglected. The opposite was true. Governments, standards bodies, researchers and frontier artificial intelligence laboratories were already investing substantial intellectual and technical resources into understanding and managing the risks associated with increasingly capable systems.

However, the research also revealed a boundary.

Most of the safety architecture I was examining began with artificial intelligence itself.

My research had arrived at the problem from somewhere different. I had started with the human. That investigation had led me to labour. Labour had led me to technology. Technology had led me to infrastructure. Infrastructure had led me to execution. Execution had led me to authority. Authority had eventually led me to artificial intelligence safety.

That sequence changed the object I was looking at.

5. Research Implications

An artificial intelligence model cannot move money merely because it is intelligent. It cannot hire someone merely because it can reason. It cannot operate a factory merely because it can generate a plan. It cannot alter a consequential institutional system merely because it possesses a dangerous objective.

Something else has to exist.

The model requires access to systems through which it can act. It requires permissions that determine what it is allowed to do. It requires interfaces through which instructions can pass. It requires resources upon which those instructions can operate. It requires infrastructure capable of translating cognition into action.

This brought the investigation back to the Infrastructure Condition.

The model matters, but so do the permissions surrounding it. The APIs connected to the model matter. The identity systems authenticating it matter. The financial systems accessible to it matter. The software receiving its instructions matters. The physical systems responding to it matter. The constraints embedded around execution matter. The humans capable of intervening matter. The mechanisms capable of stopping execution matter.

Most importantly, the infrastructure carrying decisions into consequence matters.

This produces a distinction between artificial intelligence safety and Execution Authority Governance.

Artificial intelligence safety is substantially concerned with how increasingly capable artificial intelligence systems can be made safer. My research is asking a related but different question: how do we govern what increasingly capable systems are actually authorised and able to execute?

This distinction matters because intelligence creates capability, while execution creates consequence.

A system may possess significant capability without possessing significant authority. Conversely, a system with limited intelligence can produce substantial consequences if it has been granted extensive execution authority.

The governance problem therefore cannot end at the model.

It must extend to the conditions under which the model is permitted to act, the systems to which it is connected, the resources available to it, the constraints imposed upon execution and the mechanisms through which human authority can intervene.

If execution authority is migrating from humans to systems, then the safety question becomes inseparable from the governance of that migration.

6. Conclusion

The experiment began as a search for safety nets.

Those safety nets existed. They were emerging through regulation, risk-management frameworks, technical safety research, capability evaluations, alignment techniques, monitoring, safeguards and organisational controls.

However, the investigation exposed a wider governance problem.

Controlling what an artificial intelligence system thinks, generates or intends is not identical to controlling what it can execute.

The consequential boundary therefore does not sit only at the model. It also sits at the point where cognition encounters infrastructure.

The question was no longer simply whether artificial intelligence could be made safe. It was becoming a question of authority.

The research now had to ask who or what was permitted to act, under what conditions that authority could be exercised, through which infrastructure execution could occur, what constraints surrounded that execution, whose consent authorised it and who retained the capacity to stop execution once it had begun.

Those questions would take the research beyond artificial intelligence safety and towards the governance of Execution Authority.

7. References

Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.

Russell, S. (2021). “Artificial Intelligence and the Problem of Control.” In Human-Like Machine Intelligence. Oxford University Press.

European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union.

Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, NIST AI 100-1.

National Institute of Standards and Technology. NIST AI Resource Center (AIRC). Trustworthy and Responsible AI Resource Center.

Anthropic. Responsible Scaling Policy.

Anthropic. Frontier Safety Roadmap.

Anthropic. (2024). Sabotage Evaluations for Frontier Models.

Bowman, S. R., Wagner, M., Roger, F. & Karnofsky, H. (2025). Anthropic’s Pilot Sabotage Risk Report.

Anthropic. (2022). Constitutional AI: Harmlessness from AI Feedback.