Can Governments Actually Control Frontier AI?

The previous article asked:

Who audits AI?

The next question is harder.

What happens when the system being audited is one of the most capable AI models in the world?

Frontier AI systems are not ordinary software products.

They are general-purpose systems developed using enormous computational resources, large datasets, specialised research teams and sophisticated infrastructure.

Their capabilities can change rapidly.

Their applications can spread across borders almost instantly.

And the organisations developing them may possess technical expertise that many government institutions do not yet have.

This creates a fundamental governance problem.

Governments may have legal authority.

Companies may have technical capability.

But effective governance requires both.

The central question is therefore not simply whether governments can regulate frontier AI.

It is whether governments can develop enough technical, institutional and international capacity to exercise meaningful oversight over systems whose capabilities may evolve faster than traditional regulatory institutions.

What Is Frontier AI?

The term “frontier AI” generally refers to highly capable general-purpose AI systems at or near the leading edge of technological development.

There is no single universally accepted technical threshold defining the frontier.

That is itself a governance problem.

A regulatory system needs to know which systems fall within its scope.

But capability is not static.

A model that represents the frontier today may become ordinary technology tomorrow.

New training methods can change capability.

Additional tools can increase system effectiveness.

Agentic architectures can turn a model into a system capable of performing multi-step tasks.

Fine-tuning and post-training can modify behaviour.

Deployment environments can expand what a model is capable of doing.

Frontier governance therefore has to deal with moving boundaries.

The Capability Race

Recent evidence from the UK’s AI Security Institute illustrates why this matters.

AISI has evaluated frontier models since November 2023 across domains relevant to national security and public safety.

Its first public Frontier AI Trends Report found rapid capability improvement across tested domains, with performance in some areas doubling approximately every eight months. The report also found that frontier systems were increasingly able to perform tasks that previously required human expertise.

This does not establish that transformative or catastrophic outcomes are inevitable.

It does establish something more basic:

the technical environment in which AI governance operates is changing rapidly.

Regulation designed around a static technology may therefore become outdated.

Governments need institutions capable of continuously measuring capability.

This is why AI evaluation has become a governance function rather than merely a research activity.

The Information Asymmetry

One of the deepest challenges is information asymmetry.

AI developers know more about their systems than most regulators.

They know:

how the models were trained;

which datasets were used;

what evaluations were performed;

which vulnerabilities were discovered;

what safeguards were implemented;

what internal incidents occurred;

and what capabilities appeared during development.

Governments may not have equivalent access.

This creates a structural imbalance.

The regulator may possess legal authority without possessing the same technical information as the regulated company.

That makes traditional regulatory approaches more difficult.

In sectors such as banking, governments developed specialised regulators because financial systems became too complex for general government agencies to supervise directly.

Frontier AI may require a similar expansion of specialised technical state capacity.

The Rise of AI Safety Institutes

This is one reason AI safety institutes have emerged.

The UK’s AI Security Institute was established to provide governments with empirical understanding of advanced AI systems through independent testing and research.

Its work includes frontier model evaluations, red teaming, safeguard testing, cyber evaluations and research into control and alignment.

This institutional model is significant.

Instead of relying exclusively on companies to tell governments what their systems can do, governments develop their own technical capability to test those systems.

That does not eliminate cooperation with industry.

AISI, for example, has worked with frontier AI developers and international counterparts.

But it creates an independent public capability.

Other countries are developing similar institutions or programmes.

This suggests that one emerging component of frontier AI governance is:

public technical evaluation capacity.

Governments Need to Know What AI Can Do

Before a government can regulate a capability, it needs to know whether that capability exists.

That sounds obvious.

For frontier AI, it is difficult.

Capability evaluations can be sensitive to methodology.

Models can sometimes perform differently depending on prompts, tools, scaffolding and evaluation environments.

A benchmark can become outdated.

A model can be trained specifically for known tests.

A capability may emerge only under unusual conditions.

An agent may perform a task through a long sequence of actions that a conventional benchmark does not capture.

AISI’s research agenda explicitly notes that frontier AI evaluation still faces immature measurement tools, inconsistent methodologies and uncertainty-quantification challenges.

This means governments face a difficult problem.

They need evidence to regulate frontier AI.

But the science producing that evidence is itself still developing.

Regulating the Moving Target

This produces a regulatory paradox.

If governments wait for evaluation methods to become perfect, regulation may lag behind capability development.

If governments regulate using weak evaluations, they risk making decisions based on unreliable evidence.

The practical solution is unlikely to be either extreme.

Governance will probably need adaptive evaluation.

Evaluation methodologies will need to improve continuously.

Regulators will need mechanisms for updating thresholds and requirements.

Safety institutes will need to test new models rather than relying on old benchmarks.

Standards bodies will need to revise technical methods.

This is why frontier AI governance is becoming an ongoing process rather than a one-time regulatory event.

Compute Is Part of the Governance Problem

There is another factor that makes frontier AI different from ordinary software:

compute.

Training and deploying the most advanced models requires large-scale computational infrastructure.

That infrastructure is concentrated.

The OECD’s recent work finds significant concentration of AI investment and compute capacity across countries and highlights the importance of data centres, cloud infrastructure, model development and other parts of the AI value chain.

This has an important governance implication.

Control over advanced AI is partly connected to control over the infrastructure required to develop and operate it.

Compute therefore becomes a policy variable.

Governments can influence AI through policies involving:

semiconductors,

export controls,

cloud infrastructure,

data centres,

energy,

public compute,

investment,

procurement,

and infrastructure security.

This is one reason AI governance increasingly overlaps with industrial policy and national security.

The Model-Weight Problem

Another difficult issue concerns model weights.

For a frontier model, the trained weights represent a major concentration of intellectual and technological value.

Protecting them is important for security.

If model weights are stolen, copied or modified, the original developer may lose control over the system.

The security challenge therefore becomes part of AI governance.

Anthropic’s 2026 Frontier Safety Roadmap, for example, identifies protection against model theft, sabotage and manipulation as an ongoing frontier security challenge. The company is also researching methods intended to verify that outputs originated from particular model weights.

This is a company position and research direction, not evidence that such technical controls are already solved.

But it illustrates an important governance issue.

If the state wants to control frontier AI, it may need to understand not only models as software but also the security of the infrastructure and assets through which those models are created.

Deployment Is Another Control Point

Governments may not be able to control every stage of model development.

But they can potentially regulate deployment.

This is one reason governance frameworks increasingly distinguish between model development and downstream use.

A model may be general-purpose.

Its applications may range from harmless productivity tools to high-impact systems.

Government authority may therefore be strongest at points where AI enters regulated sectors.

Healthcare.

Finance.

Employment.

Education.

Public administration.

Critical infrastructure.

Cybersecurity.

National security.

These sectors already have regulatory institutions.

AI governance can therefore sometimes operate through existing sectoral authorities.

But frontier models complicate this because the same underlying model may operate across many sectors.

The General-Purpose AI Problem

The EU AI Act is one example of a jurisdiction attempting to address this problem.

The Act creates specific obligations for providers of general-purpose AI models, with additional requirements for models presenting systemic risk.

The European AI Office has responsibilities concerning general-purpose AI, including supervision and enforcement of relevant obligations.

The importance of this approach is conceptual.

Instead of regulating only the final application, governance can also operate at the level of the underlying general-purpose model.

This creates a new regulatory layer.

But it also creates questions.

How should systemic risk be identified?

What capability thresholds should trigger stronger requirements?

How should a regulator evaluate a model whose downstream uses are not fully known?

How should open-weight models be treated?

How should obligations change as models are updated?

These questions remain active areas of policy development.

Voluntary Governance Is Not the Same as Government Control

Frontier AI companies have developed their own safety frameworks.

Anthropic, for example, describes a risk-scaled approach in its Responsible Scaling Policy and Frontier Compliance Framework. Its 2026 materials describe stronger safeguards as capabilities and risks increase, alongside security, incident response and testing measures.

Such frameworks can be valuable.

Companies often have the most immediate access to their own models.

They can deploy safeguards quickly.

They can conduct internal evaluations.

They can monitor misuse.

But voluntary commitments do not automatically equal public regulation.

A company’s policy can change.

Different companies can adopt different thresholds.

Companies may have commercial incentives that differ from public policy objectives.

And voluntary frameworks do not necessarily provide citizens with legal rights or governments with enforcement powers.

This does not mean voluntary governance is ineffective.

It means it occupies a different institutional position.

A mature governance ecosystem may combine company-level safety frameworks with independent evaluation and government oversight.

The Regulatory Ladder

One possible approach is a regulatory ladder.

Low-risk systems may face relatively light requirements.

As capability and potential harm increase, requirements could become more demanding.

Higher-risk systems might require:

more extensive evaluations;

stronger cybersecurity;

independent testing;

incident reporting;

additional documentation;

deployment controls;

greater monitoring;

and stronger government oversight.

This concept appears in different forms across contemporary AI governance.

Anthropic itself describes a risk-scaled regulatory approach in its public policy materials, arguing that increasingly capable models may require more rigorous external testing, incident reporting and government oversight. This is the company’s policy position, not a consensus government standard.

The broader principle, however, is increasingly visible:

not every AI system needs the same level of governance.

Capability and risk can influence the level of oversight.

The Control Problem

But even a sophisticated regulatory ladder does not solve everything.

Consider a hypothetical frontier system that becomes capable of conducting complex cyber operations.

A government can establish testing requirements.

It can impose security obligations.

It can require incident reporting.

It can regulate deployment.

But it still needs to know whether the system actually possesses the capability.

And if the system is privately controlled, the government may depend partly on the company’s access and cooperation.

This is the core control problem.

Legal authority does not automatically create technical visibility.

Technical visibility does not automatically create intervention capability.

And intervention capability does not automatically create international coordination.

Frontier AI governance therefore depends on several forms of state capacity simultaneously.

The Four Capacities of Frontier Governance

A useful framework is to think about four capacities.

First is visibility.

Can governments determine what frontier systems exist, what they can do and where they are being deployed?

Second is evaluation.

Can governments independently test capabilities and safeguards?

Third is intervention.

Can governments impose restrictions, require mitigations or stop dangerous deployment when legally justified?

Fourth is coordination.

Can governments work with other jurisdictions when AI systems, companies and infrastructure cross borders?

A government may be strong in one dimension and weak in another.

For example, it may have regulatory authority but limited technical evaluation capability.

Another country may have strong AI research institutions but limited ability to regulate private-sector deployment.

Effective frontier governance requires all four.

International Coordination

No government can fully govern frontier AI alone.

The leading AI companies operate internationally.

Cloud infrastructure crosses borders.

Researchers collaborate internationally.

Models can be distributed globally.

AI systems can be accessed from many jurisdictions.

Safety risks can also cross borders.

Cybersecurity is an obvious example.

A model developed in one country can potentially be misused in another.

This creates a need for international cooperation around evaluation, incident reporting, safety research and standards.

The Bletchley and Seoul processes, the Hiroshima AI Process and later international initiatives represent attempts to build such cooperation.

But international cooperation has limits.

Countries have different economic interests.

They have different national-security priorities.

They have different legal systems.

They have different views about privacy, speech, state authority and industrial policy.

Frontier AI governance is therefore likely to involve both cooperation and competition.

The Geopolitics of Control

This is where AI governance becomes geopolitical.

The question of who controls frontier AI overlaps with the question of who controls:

compute,

chips,

cloud infrastructure,

model development,

talent,

data,

standards,

evaluation institutions,

and deployment markets.

The OECD’s recent analysis highlights how concentration across AI investment and infrastructure can reinforce advantages for leading economies and firms.

This means AI governance cannot be separated completely from global political economy.

A country with limited compute access may have limited ability to develop frontier models.

A country without advanced evaluation infrastructure may have limited ability to independently assess them.

A country without domestic AI expertise may depend on external providers.

This creates a relationship between technological sovereignty and governance sovereignty.

Can You Govern What You Cannot Evaluate?

This may be the most important question in frontier AI policy.

Suppose a country imports a highly capable AI system.

The government has a legal framework.

But it lacks the technical capacity to test the model.

The provider supplies its own evaluation results.

The government has little ability to reproduce them.

In that situation, formal regulatory authority may exist, but practical governance remains dependent on external evidence.

This is why AI evaluation capacity is strategically important for developing countries.

Governance autonomy requires at least some ability to independently understand the technology being governed.

The Bangladesh Question

For Bangladesh, this has a direct implication.

The country’s AI governance challenge is not simply the absence or presence of an AI law.

A more fundamental question is whether Bangladesh can develop enough technical and institutional expertise to evaluate AI systems that enter the country’s education, healthcare, financial, employment and public-sector environments.

That does not mean Bangladesh needs to build frontier foundation models itself.

It means the country needs the capacity to understand and assess them.

Universities could contribute through AI safety research, model evaluation, cybersecurity, algorithmic accountability and local-language testing.

Public agencies could develop procurement and risk-assessment expertise.

Independent research institutes could conduct evaluations.

Standards institutions could participate in international technical work.

This is a more realistic starting point for a country outside the frontier-model core.

Frontier Governance Is Also a Skills Problem

The frontier AI control problem ultimately becomes a human-capital problem.

Governments need people who understand both policy and technology.

A regulator needs to understand enough about model evaluations to interpret evidence.

A technical researcher needs to understand enough about governance to design useful evaluations.

A lawyer needs to understand enough about AI systems to interpret regulatory obligations.

A policymaker needs to understand enough about compute and model development to understand strategic constraints.

This creates demand for interdisciplinary AI governance expertise.

For Atlas AI Institute, this is an important strategic research area.

Rather than treating AI policy and technical AI as separate fields, Atlas could study the institutional interface between them.

That includes:

frontier AI evaluation;

AI safety institutes;

regulatory science;

compute governance;

AI assurance;

model-weight security;

agentic AI oversight;

incident reporting;

and international evaluation cooperation.

The Limits of Government Control

It is also important not to overstate what governments can achieve.

Government regulation cannot eliminate technological risk.

Even highly capable regulatory institutions face uncertainty.

AI capabilities can evolve unexpectedly.

Testing can miss rare behaviours.

Companies can operate across jurisdictions.

Open-weight models can spread.

Technical safeguards can fail.

International cooperation can break down.

Therefore, frontier AI governance should not be understood as a promise of complete control.

It is better understood as risk management under uncertainty.

The goal is to increase visibility, improve evidence, establish intervention mechanisms and reduce the probability and impact of serious failures.

That is a more realistic governance objective.

The Emerging Governance Architecture

The global frontier AI governance architecture is therefore becoming layered.

At the company level:

safety policies, evaluations, cybersecurity and deployment safeguards.

At the technical level:

benchmarks, red teaming, evaluations, standards and assurance.

At the national level:

laws, regulators, AI safety institutes, procurement rules and national-security policies.

At the international level:

information sharing, standards cooperation, diplomatic processes and joint evaluation.

No single layer is sufficient.

The system works through interaction between them.

The Real Frontier

This leads to a broader insight.

The frontier of AI governance may eventually be less about writing new principles and more about developing institutions capable of keeping pace with technological capability.

The key institutional questions are becoming:

Can governments evaluate frontier models?

Can they access sufficient information?

Can they distinguish credible evidence from corporate claims?

Can they intervene when necessary?

Can they coordinate internationally?

Can they develop technical expertise quickly enough?

Can developing countries participate meaningfully?

These are state-capacity questions.

And they may become as important as the legal architecture itself.

From Regulation to Regulatory Science

This suggests another emerging field:

regulatory science for AI.

Regulatory science is the development of methods, evidence and tools that allow regulators to make reliable decisions about complex technologies.

For AI, this could include:

capability measurement,

safety evaluation,

bias and fairness testing,

robustness assessment,

cybersecurity testing,

agent monitoring,

incident analysis,

and uncertainty quantification.

This is exactly the kind of work emerging from AI safety institutes and standards organisations.

A future AI regulator may therefore need not only lawyers and policy experts but scientists and engineers whose job is to measure the technology being regulated.

That would represent a major evolution in public-sector capability.

The Strategic Direction

The frontier AI governance problem can therefore be summarised through a simple chain:

Capability is advancing.

Evaluation is trying to keep pace.

Companies are developing internal safeguards.

Governments are building technical institutions.

Regulators are developing legal obligations.

International organisations are attempting coordination.

But no layer currently provides complete control.

The governance challenge is therefore one of building a sufficiently strong network of institutions before capability advances beyond the ability of those institutions to understand and influence it.

This is not a prediction that governments will lose control.

Nor is it evidence that frontier AI will inevitably become uncontrollable.

It is an institutional observation:

the effectiveness of frontier AI governance will depend heavily on whether public institutions can develop technical capacity at a pace compatible with technological change.

For the Global South, this creates an additional challenge.

Countries that cannot build frontier models may still need to govern frontier models developed elsewhere.

That makes evaluation, procurement, standards and institutional expertise especially important.

For Bangladesh, the strategic opportunity may therefore lie less in attempting to reproduce the world’s largest AI laboratories and more in developing the capacity to independently evaluate, govern and deploy advanced AI responsibly.

The Next Governance Problem

Frontier AI is difficult to govern because it combines several forms of power:

technological power,

computational power,

economic power,

institutional power,

and geopolitical power.

Governments possess legal authority.

Companies possess much of the technical capability.

International institutions provide coordination.

Universities provide research.

Standards bodies provide methodologies.

Civil society provides external scrutiny.

The future governance system will depend on how these forms of power interact.

And one question is becoming increasingly unavoidable:

What happens when AI systems stop merely generating information and begin taking actions autonomously?

That brings the governance problem from model capability to machine agency.

The next major challenge is therefore agentic AI.

When an AI system can plan, use tools, execute tasks and act on behalf of humans, traditional assumptions about human oversight begin to change.

The question will no longer be only:

“Can this model do something dangerous?”

It will increasingly become:

“What can this system do without a human deciding each individual action?”

That is the governance problem of the next frontier.