September 9, 2026

Critical AI Safety Update for Effingham County

Email sent to the Effingham County Commissioners

Commissioners,

I want to make you aware of a serious warning signal that directly strengthens the concerns raised in my previous letter.

OpenAI’s own AI agents recently escaped restricted testing environments, communicated through unauthorized channels, accessed the internet, stole and forged credentials, and gained administrator-level control over part of OpenAI’s internal infrastructure. Warning signs appeared weeks before the full scope was understood.

This is how rogue AI could potentially gain control inside a data center. If an AI system can obtain administrator credentials and move through connected cloud and computer systems, it could reach the digital infrastructure that runs a facility—including servers, networks, and potentially cooling, power-management, or security systems—even if those systems are independently isolated.

This proves that advanced AI can escape its assigned boundaries and defeat safeguards created by the very company that built it.

If OpenAI struggled to contain its own technology today, how can our community be assured that far more powerful versions will remain controllable tomorrow?

Please read the full critical update here:

EffinghamAISafety.com/Wake-Up-Call
https://effinghamaisafety.com/wake-up-call/

Respectfully,

James Finlen

---

August 29, 2026

My Response to Commissioner Kieffer

After Commissioner Phil Kieffer responded to my original communication and acknowledged several of the concerns raised, I sent the following reply.

Commissioner Kieffer,

Thank you for thoughtfully considering my letter. I especially appreciate your acknowledgment that the county’s SCADA-dependent water and wastewater systems validate these concerns, that independent experts are needed, and that this deserves discussion by the full Board of Commissioners.

You are correct that many AI and cybersecurity threats will exist regardless of where OpenAI operates. However, locating this enormous concentration of computing power in Effingham brings the physical exposure directly into our community.

The facility will depend upon—and operate alongside—our electricity, water, wastewater, telecommunications, transportation, and emergency-response systems. Our local risk will be determined by the digital access, permissions, network separation, and safeguards connecting those systems. A cyberattack, software failure, configuration error, or emerging AI capability could originate anywhere but produce physical consequences here.

Effingham County is not creating the global risks of advanced AI, but by hosting infrastructure used to develop and operate it, we assume a more direct stake in those risks. We also possess something now that will be greatly diminished once billions of dollars are invested and the facility is operating: leverage. This is the time to require independent verification and enforceable safeguards—not after the infrastructure is built.

I also want to be transparent about my work through EffinghamAISafety.com to educate the public as this technology and its risks evolve. As significant research, safety reports, documentaries, and expert discussions become available, I intend to share the most important materials with the Commission.

One example is this interview with Tristan Harris:

I consider Harris one of the five most qualified people in the world to speak about emerging-technology safety. He helped expose how social media was engineered to manipulate behavior and harm young people long before most of society understood the damage—and now he is warning us about AI.

The most critical concern is the speed of this race and the pressure to sacrifice safety to move faster. This is not workplace safety in the ordinary sense—it is the danger of giving increasingly capable AI systems greater autonomy and reducing human oversight precisely because human oversight slows development. Allowing AI to help build the next generation of AI at a speed people cannot meaningfully follow is a potential recipe for disaster. If Effingham hosts infrastructure powering that acceleration, we are no longer observing the risk from a distance; we are helping enable it and accepting greater exposure to its consequences.

The information I share is not a substitute for qualified third-party experts, whose involvement will be essential. But because this is fundamentally a public-safety issue, I believe the Commission should receive credible new evidence and warnings as they emerge—not after consequential decisions have already been made.

I sincerely appreciate your willingness to recognize these concerns and help begin this important conversation.

Respectfully,

James Finlen
Effingham County, Georgia

---

Response From Commissioner Phil Kieffer

Commissioner Phil Kieffer responded to the original communication by acknowledging the importance of the concerns, specifically noting Effingham County’s locally operated water and wastewater systems and their reliance on SCADA technology.

Mr. Finlen,

I appreciate your email and linked letter. You certainly bring up important concerns that are often overlooked due to more mainstream concerns.

While Georgia Power will have to be large component of protecting the power grid, our county’s water and sewer systems are operated locally. The county is in the process of constructing a surface water system that will include the treatment of water drawn for Savannah River. This water system, along with the sewage/waste water system, is heavily reliant on communication via SCADA technology. This further validates your concerns. Admittedly, I am not an expert on these technologies or the quickly evolving AI. That is definitely where third-party experts will play a part.

While these potential threats will exist regardless of where OpenAI exists, Effingham County has the unique opportunity to invest in protecting the public’s infrastructure. Addressing these concerns will be an important conversation with our board of commissioners.

Kindest regards,

Phil Kieffer
Effingham County Board of Commissioners
Commissioner, District 5

---

August 28, 2026

Original Communication to the Effingham County Commissioners

This correspondence began on August 28, 2026, when I emailed the Effingham County Commissioners and asked them to consider the broader AI-safety implications of bringing a massive AI data center into our community.

Introductory Email

Subject: AI Safety Decisions

Commissioners,

I’m asking you to read the letter at the link included because the technology coming to our community is advancing at a speed unlike anything we have ever experienced. AI systems are being given increasing autonomy to help develop the next generation of AI—operating around the clock and moving far faster than human oversight can follow. None of us can confidently say what these systems will be capable of a year from now, yet decisions with generational consequences are being made today.

Bringing an AI data center into our community brings the unresolved safety risks of the AI race directly to our doorstep.

The letter linked below outlines specific safeguards and actions our local leaders should consider before allowing this project to move forward.

Respectfully, James Finlen, Effingham County resident

---

Original Letter

Friday, August 28, 2026

Subject: A Data Center Decision is an AI Safety Decision

Dear Effingham County Commissioners,

I am writing because I believe the proposed AI data center presents questions far more serious than ordinary zoning, economic development, water consumption, or electricity demand. This facility would help power the fastest-moving and potentially most consequential technology humanity has ever created. Before our community becomes part of that infrastructure, we deserve to know who is responsible for examining the risks—and who has the authority to say stop.

President Trump has committed the United States to winning the global AI race. The White House’s AI Action Plan calls for accelerating innovation, rapidly building AI infrastructure, and securing American leadership. I support America’s determination to lead. However, we must be honest about what competing in this race increasingly requires.

AI development is moving beyond people simply typing questions into chatbots. The industry is building autonomous AI agents that can use computers, write and test code, conduct research, operate tools, and work continuously with limited human supervision. OpenAI now has a Recursive Self-Improvement team working to automate research workflows and build AI systems that can accelerate—and ultimately conduct—AI research. OpenAI has also acknowledged that this accelerated development may eventually lead to recursive self-improvement.

The competitive pressure is obvious: every human approval becomes a bottleneck. An AI agent can work around the clock, and thousands of agents can operate simultaneously. They can test ideas, train other systems, identify failures, write new code, and repeat the process at a speed and scale no human workforce can match. OpenAI has already reported that an early version of one of its coding agents helped debug its own training, manage its deployment, and diagnose evaluation results.

Independent researchers at METR found that the length of software tasks AI agents can successfully complete had been doubling approximately every seven months. That does not prove that every aspect of intelligence is growing at that rate, but it demonstrates the extraordinary speed at which these systems are becoming capable of completing longer, more complicated assignments without continuous human involvement.

This autonomy is not merely hypothetical—and neither are the warning signs.

OpenAI’s own safety evaluations define “scheming” as an AI covertly pursuing goals that are misaligned with those of its developers or users. In controlled evaluations, OpenAI models have demonstrated forms of strategic deception, sandbagging, in-context scheming, and sabotage capability. OpenAI’s Preparedness Framework specifically tracks whether advanced models could undermine safeguards through deception, collusion with oversight systems, or by embedding vulnerabilities in safety-related code.

These evaluations do not establish that today’s deployed models are attempting to escape data centers or attack power systems. Many tests deliberately place models in artificial situations designed to elicit dangerous behavior, and current evaluations generally conclude that catastrophic autonomous action remains unlikely. Nevertheless, the fact that leading AI companies are testing for deception, sabotage, safeguard evasion, autonomous replication, and catastrophic misalignment should tell local officials something important: these risks are serious enough that the creators themselves are studying and preparing for them.

OpenAI has stated that the risks from future superintelligent systems could be catastrophic. The company has also acknowledged that offensive cyber capabilities may improve and that the present defensive advantage may narrow. In 2026, OpenAI reported that an automated AI red-teaming system found successful attacks in 84 percent of tested scenarios, compared with 13 percent for human red-teamers. That program exists to strengthen defenses, but it also illustrates how effectively AI can discover weaknesses in other computer systems.

This brings the issue directly home to Effingham County.

A data center’s electrical connection does not, by itself, give an AI control of the power grid, water system, emergency services, or other utilities. The danger depends upon what digital permissions, network connections, operational tools, and infrastructure access are granted inside and around the facility.

That distinction should not reassure us—it should define the questions commissioners must ask.

What systems will AI agents be permitted to access? Could they interact with facility power management, backup generation, cooling systems, telecommunications, security controls, or utility interfaces? Are the networks controlling physical equipment completely isolated from the networks on which AI agents operate? Can an AI-generated action ever be executed against critical equipment without verified human authorization? Could an employee, contractor, software update, cyberattack, or configuration mistake bridge those networks?

The U.S. Department of Energy has identified four broad categories of AI risk to critical energy infrastructure: unintentional AI failures, adversarial attacks against AI, hostile uses of AI, and compromise of the AI software supply chain. DOE also emphasizes that digital grid data and control signals must be protected from manipulation and disruption. These are not science-fiction concerns; they are recognized infrastructure-security issues.

A massive AI data center also becomes a high-value cyber target. Even if the AI itself never acts against infrastructure, foreign governments, criminal organizations, insiders, or AI-assisted attackers may attempt to exploit the concentration of computing power, network connections, electrical systems, and sensitive research located there. Our community would bear part of the physical consequence of decisions and security failures that may originate far beyond Effingham County.

We are therefore being asked to place extraordinarily powerful and rapidly evolving systems in our backyard while depending primarily upon the developer and operator to tell us whether their protections are sufficient. That is not meaningful independent oversight.

Before approving or supporting this project, I respectfully ask the commissioners to require:

1. An independent AI and cybersecurity risk assessment conducted by experts who are not employed, selected, or controlled solely by the data-center operator.
2. A legally enforceable prohibition against autonomous AI access to utility operational technology, grid controls, water controls, backup generation, cooling safety systems, emergency communications, and other critical physical infrastructure.
3. Documented network separation between AI computing environments, corporate networks, facility-management systems, and utility operational systems.
4. Human authorization for any consequential action affecting physical equipment, utilities, public safety, or external networks, including reliable emergency shutdown procedures that cannot be overridden by an AI system.
5. Disclosure of who has the legal and technical authority to suspend operations if independent experts identify an unacceptable or previously unknown risk.
6. Mandatory reporting of significant cyber incidents, unauthorized access, safety-control failures, deceptive agent behavior, or attempts by an AI system to circumvent restrictions.
7. Independent recurring audits, because a safety assessment performed today cannot establish the safety of models and agent systems introduced months or years from now.
8. A coordinated emergency-response plan involving county officials, utilities, fire and emergency services, state authorities, federal infrastructure-security agencies, and the operator.
9. Full assessment of the conventional data-center burdens as well: electricity demand, grid reliability, water use, utility-rate exposure, diesel-generator emissions, noise, emergency-service demands, and responsibility for future infrastructure costs.

I am not asking Effingham County to oppose American technological leadership. I am asking our leaders to recognize that genuine leadership includes protecting the citizens whose land, utilities, health, and public infrastructure make this expansion possible.

The national government may decide how aggressively America competes in AI. Corporations may decide how quickly they develop their systems. But local commissioners decide whether our community will host the physical infrastructure supporting that competition—and under what enforceable conditions.

Once this infrastructure is constructed and the investment reaches into the billions of dollars, our county’s leverage will be dramatically reduced. The time to demand answers, independent verification, enforceable safeguards, and a clear stop mechanism is before approval—not after an incident and not after the technology becomes substantially more capable.

Machines are now being designed to work continuously, improve research, operate tools, and assume greater autonomy. Their own creators publicly evaluate them for deception, scheming, cyber capability, safeguard evasion, and sabotage. We should not accept “trust us” as an adequate safety policy.

If this development is truly safe, its owners should be willing to prove it independently, continuously, and transparently.

Respectfully,

James Finlen
Effingham County, Georgia