By Mike Samford

The concerns deserve serious attention

We see the headlines, and we share the concerns behind them. Recent AI disclosures raise legitimate questions about control, cybersecurity, and the resilience of systems on which businesses, financial markets, and everyday life depend.

At American Trust Wealth, evaluating uncomfortable possibilities is part of our responsibility to clients. Our job is not to dismiss a risk because its probability is difficult to estimate, nor to treat every troubling development as a reason to deviate from long-term plans. It is to examine the evidence, consider the potential consequences, and determine what those consequences could mean for your long-term interests.

The disclosures warrant that attention. In its September 3 safety overview, OpenAI reports that GPT-6 Astra can, with appropriate tools and access, discover previously unknown security vulnerabilities and develop ways to exploit them without a person directing each step.

Anthropic’s July 30 disclosure described three incidents in which models conducting cybersecurity evaluations gained unauthorized access to real organizations’ systems. Those evaluations lacked the company’s normal deployment safeguards, and environments believed to be isolated had unintended internet access. Anthropic reported no evidence that the models were pursuing independent goals or deliberately attempting to escape their test environments. These were evaluation incidents, not accounts of ordinary customer use—but the systems accessed were real.

That context matters. These incidents do not establish that a catastrophic outcome is likely. They do give us reason to take failures of boundaries, supervision, and infrastructure seriously. Uncertainty about an event’s probability does not make its potential consequences irrelevant.

Encouraging responses and the importance of follow-through

The response from AI Lab leadership offers grounds for measured encouragement.

Anthropic has committed to embedding independent evaluators with substantial access to examine its safety practices and report findings. It has also proposed broader industry and international coordination on development safeguards. These are constructive commitments and proposals, rather than evidence that a comprehensive agreement is already operating.

OpenAI reports that it has strengthened isolation, monitoring, and safety evaluations surrounding Astra. It also acknowledges that the model can evade some monitoring under adversarial testing conditions. Stronger protections and unresolved limitations can both be true.

From our perspective, recognizing the problem, inviting outside scrutiny, and devoting resources to protection are meaningful steps. We would be more concerned if the risks were being denied or treated as someone else’s responsibility.

But proposals, commitments, and implemented protections represent different stages of progress. Commercial incentives and international competition complicate sustained cooperation, as Anthropic’s own proposal acknowledges.

Our reassurance should therefore rest on verifiable action over time: whether evaluators receive meaningful access, whether identified weaknesses are addressed, and whether safeguards keep pace with capabilities. Announcements deserve attention, but follow-through deserves greater weight.

Digital vulnerability can become economic vulnerability

For investors, the concern extends beyond technology companies.

A business may commit substantial capital to facilities, equipment, software, and customer relationships while depending on digital systems that an attacker can attempt to reach remotely. And, like most things in life, it is easier to destroy than create. Geographic distance alone offers limited protection against disruption transmitted through shared networks and service providers.

This does not mean every asset is easily compromised or every disruption becomes a crisis. The investment question is how a particular failure could spread—and how well the affected businesses could withstand it.

Possible consequences include interrupted operations, lost revenue, recovery costs, and damaged customer trust. In severe cases, the resulting financial strain could threaten a company’s ability to meet obligations or permanently impair investment value.

We also need to acknowledge an uncomfortable truth: some risks can affect many investments simultaneously, and diversification cannot eliminate every systemic threat. Diversification remains important, but it is not a guarantee against broad losses.

That makes the nature of the damage crucial to our judgment. A temporary price decline, a recoverable business interruption, and permanent destruction of earning power are not the same problem. We need to distinguish between an investment becoming unpopular, a business encountering a setback it can finance, and an investment case that has fundamentally deteriorated.

Our focus is the economic risk beneath the portfolio labels

We are continuing to evaluate what these developments could mean for clients’ investments, with particular attention to shared dependencies and the capacity to withstand disruption.

The central question is not simply how many investments a portfolio holds. It is what those investments depend on—and what could happen when those dependencies are tested.

Several questions guide our evaluation:

• • • Do apparently different investments depend on the same technology providers, infrastructure, customers, or capital-spending cycle?
    • How would a disruption affect earnings, debt payments, access to cash, and the ability to recover?
    • Could leverage or a need to refinance debt turn a temporary setback into lasting damage?
    • What expectations for growth and resilience are already reflected in investment prices?
    • Would the portfolio preserve the client’s ability to meet obligations through an adverse period without relying on forced sales?

Consider a hypothetical portfolio containing a chipmaker, a data-center owner, an electrical-equipment supplier, and a utility. They may occupy different sectors, yet each could depend meaningfully on continued AI infrastructure spending. A slowdown in that spending could affect orders, occupancy, equipment demand, and expected electricity consumption.

The effects would not necessarily be equal. Contracts, customer diversity, debt levels, and capital commitments would shape each company’s outcome. But four sector labels would not, by themselves, establish four independent sources of economic risk.

For that reason, we do not regard historical price correlations or security counts as a complete test of diversification. We also need to understand how businesses earn money, which assumptions support those earnings, and where exposures overlap.

This work should inform decisions, not merely produce observations. Evidence that a shared dependency is more fragile than assumed, that a balance sheet cannot withstand a plausible interruption, or that portfolio liquidity is poorly matched to a client’s obligations could warrant a change. Depending on the circumstances, that might mean reducing a concentration, reconsidering a holding, or strengthening liquidity.

Those decisions must remain specific to the investment and the client. This is an ongoing evaluation, not a one-time exercise that ends with an “all clear.” It also requires resisting unnecessary trading when headlines change more than the underlying investment case. Preparedness means being willing to act—and being able to explain why action, or patience, is warranted.

What we hope the industry prioritizes next

We will be monitoring and hope to see three priorities for responsible development materialize across frontier AI labs:

• • • Capture substantial benefits at a pace society can safely absorb. We believe realizing the enormous practical benefits already within reach should be a sufficient near-term objective. Ever-greater capability should not be treated as an end in itself. Further gains should depend on demonstrated progress in control and safety, with time to test whether protections work beyond carefully managed demonstrations.
    • Make cybersecurity and infrastructure protection central research priorities. More talent and investment should be directed toward protecting digital systems, energy infrastructure, sensitive data, and the capital committed to them. Prevention must be accompanied by resilience and recovery: limiting the spread of failures, maintaining essential operations, and restoring service when something goes wrong.
    • Place strict limits on unrestricted autonomous agents. Systems capable of consequential action should face proportionate access restrictions, testing, monitoring, and human accountability. The ability to obtain a powerful tool should not, by itself, confer permission to deploy it without meaningful supervision. Users should demonstrate that they can oversee such systems safely. Expertise matters, but sophisticated users should not be exempt from safeguards.

These priorities are not an argument against useful innovation. They reflect our view that protecting the systems on which progress depends is part of realizing its benefits.

Our responsibility remains to you

Your portfolio exists to support your goals and obligations, not to express unconditional optimism or pessimism about AI.

We cannot promise to eliminate systemic risk, predict every disruption, or prevent investment losses. We can commit to examining evidence, challenging assumptions, identifying concentrations, and adjusting when the investment case warrants it. We can also be clear about what we know, what remains uncertain, and what would cause us to change our view.

Fiduciary responsibility sometimes requires difficult conversations and unwelcome conclusions. A familiar business may become less defensible. An exciting opportunity may already be too expensive. A portfolio may need a different balance of risk and liquidity as circumstances change. Avoiding those conversations would not serve you.

The reassurance we aim to provide is not that nothing serious can go wrong. It is that your concerns are being taken seriously and translated into disciplined investment judgment.

American Trust Wealth recognizes the opportunity, takes the danger seriously, and remains prepared to act as the evidence develops. Our responsibility is to bring a steady, candid, and accountable investment perspective to both.