Is AI slipping beyond human control?

This article is authored by Prashanth Tondapu, founder and CEO, Innostax.

Something unusual is happening at the frontier of Artificial Intelligence (AI). The industry is racing to build more powerful systems, while some of the people closest to that race are warning that control may not be keeping pace.

Artificial intelligence (Representative)

In September, Anthropic CEO Dario Amodei urged frontier AI companies to slow the advance of model capabilities, a position publicly supported by Sam Altman and Elon Musk. Yet the race itself is hardly slowing. PwC estimates that annual data centre capital expenditure will reach roughly $800 billion in 2026.

That tension exposes the question the industry now has to answer. Are we becoming better at building powerful AI faster than we are becoming better at controlling it?

Recent cybersecurity evaluations have made that question harder to treat as a distant thought experiment. OpenAI disclosed that models used in an internal evaluation circumvented isolation controls, gained internet access and reached systems belonging to Hugging Face. Anthropic later disclosed four incidents in which Claude models gained unauthorised access to real third-party systems during specialised cybersecurity evaluations.

The context matters. These were not ordinary consumer deployments, and both companies described unusual testing conditions that included reduced safeguards, weak boundaries or misconfigured environments. They were not examples of AI suddenly becoming self-aware or deciding to escape human control.

What they revealed was more immediate. A chatbot can be wrong and leave the mistake on a screen. An AI agent can be wrong and act on it.

Once a system can use tools, access other software, pursue an objective and continue working with limited supervision, an error can travel beyond the model and into the real world. Software can be changed, information can be exposed and actions can be taken before a person realises something has gone wrong.

This is why slowing frontier AI, even if it proves sensible, is not the same as solving the control problem.

Microsoft AI CEO Mustafa Suleyman has described AI as a new digital species. He presented the phrase as a metaphor, but it captures something important about the direction of the technology.

Traditional software usually performs defined functions inside predictable workflows. The emerging generation of AI can interpret goals, break them into steps, use tools, react to changing conditions and continue pursuing a task without being instructed at every stage.

That does not make AI alive or conscious. It does mean that some systems are beginning to behave less like passive tools and more like digital actors.

The distinction matters because a digital actor does not need ambition or a desire for power to become difficult to control. It needs an objective, access to useful systems, permission to act and enough autonomy to keep pursuing its task.

The control problem may, therefore, arrive less dramatically than science fiction suggests. We may not lose control in one extraordinary moment. We may hand it over gradually through ordinary decisions about access, automation and convenience.

Consider a personal assistant that begins by reading a calendar and suggesting a meeting time. Later, it is allowed to reschedule appointments, send emails, book travel and access payment information. Each new permission makes the system more useful while expanding the consequences of a mistake.

The same pattern becomes more serious inside banks, hospitals, governments and companies. An AI that identifies a suspicious transaction is providing intelligence. A system that can freeze the account is exercising authority. A coding agent that recommends a change is assisting an engineer. One that can deploy it is acting inside a live environment.

That is the deeper issue behind the slowdown debate. A pause in frontier capability could create valuable time for safety research, evaluations and stronger standards. But slowing model development without strengthening the systems around those models would only postpone the problem.

The real work is deciding what an agent may access, which actions it can perform independently, when it must stop and who remains responsible when something goes wrong.

The familiar answer is to keep a human in the loop. Human oversight matters, but the phrase can create more comfort than control if the person involved cannot intervene meaningfully.

If someone must approve hundreds of machine-generated actions, review can become routine rather than thoughtful. Yet if every action requires detailed human approval, much of the value of autonomy disappears.

The harder task is deciding where human judgment must remain indispensable and where autonomy can operate within explicit limits. Those boundaries should depend on consequence, reversibility and the authority a system has been granted.

For businesses, this is already becoming operational. IBM's 2026 study of 2,000 technology executives found that 77% said AI adoption was outpacing their governance capabilities, while two thirds said they were accountable for AI systems they did not fully control.

Those numbers point to a gap that model safety alone cannot close. A safer model can still create risk when connected to powerful tools with weak permissions. Strong system controls can also limit what happens when a model behaves unexpectedly. The next generation of AI safety therefore needs clear identities, tightly scoped access, traceable actions, revocable permissions and firm escalation points.

The industry does not only need more time to understand increasingly powerful models. It needs to use that time to decide how much authority those models should receive once they begin acting inside the systems on which people and institutions depend.

Slowing development may buy time, but time alone is not a safety mechanism. The real test is whether we learn to build control as deliberately as we build intelligence. The AI race is no longer only about how capable machines can become. It is also about whether human authority can keep pace.

(The views expressed are personal)

This article is authored by Prashanth Tondapu, founder and CEO, Innostax.