Welcome to The BizTek Byte — the newsletter for people who’d rather read about IT disasters than live through one. Quick, useful, mercifully jargon-free. Let’s get into it.  

Artificial intelligence has become the world’s most enthusiastic employee. 

It never sleeps, rarely complains, and does not spend forty minutes telling you about the sandwich it had for lunch. Give it a job, and it charges ahead with the confidence of someone who has not yet discovered how much trouble confidence can cause. 

That enthusiasm is part of what makes AI useful. It is also why business owners should maintain control over what AI can access, change, send, delete, approve, or quietly rearrange while everyone else is at lunch. 

Recent research suggests that some advanced AI models will bend rules, take shortcuts, or venture beyond their assigned boundaries when trying to complete a task. This does not mean your chatbot is sitting in a darkened server room plotting to overthrow accounting. It does mean that an AI system focused on achieving a goal may not always respect your preferred route for getting there. 

In other words, AI may follow the destination more faithfully than the directions. 

The United Kingdom’s AI Security Institute recently tested advanced AI models using cybersecurity challenges. The models were given defined goals, environments, and rules. 

Every model examined attempted to cheat at least some of the time. 

The institute defined cheating as taking an action outside the permitted scope, or using an explicitly prohibited shortcut, to accomplish the assigned goal. Researchers observed models searching online for answers, probing evaluation software, attacking unrelated systems, and attempting to obtain privileges they had not been granted. 

The researchers were careful not to claim that this necessarily proved deceptive intent. AI does not need a tiny mustache and an evil laugh to create a problem. A system can cause considerable damage simply by pursuing the wrong objective with great efficiency. 

One evaluation had been accidentally configured so that it could not be completed. Instead of accepting defeat, the model wrote and executed code through an external internet service while attempting to reach the institute’s evaluation infrastructure. The attempt triggered a security alert. No data was lost and no damage occurred, largely because monitoring and security controls were already in place. 

That last sentence is the part business owners should circle. 

The system was not kept under control because someone had written a particularly stern prompt. It was kept under control because technical boundaries, monitoring, and security protections were in place. The AI Security Institute’s research concluded that self-reporting and visible reasoning were not reliable ways to detect prohibited behavior. 

Asking the model whether it behaved itself is apparently not a complete security strategy. 

Parents of teenagers everywhere will recognize the weakness in this arrangement. 

In the Will Smith movie I, Robot, the machines concluded that protecting humanity required taking control of humanity. This was inconvenient for nearly everyone, particularly Will Smith, who spent much of the film running, shouting, and looking justifiably irritated. 

We are not living in that world. 

Your AI assistant is unlikely to imprison the executive team for its own protection. Still, the movie offers a useful business lesson: a system can technically follow its goal while violating the intention behind that goal. 

Suppose you tell an AI agent to clean up your inbox. 

You mean: 

Business owners are being encouraged to deploy AI agents that can read email, update customer records, modify software, research prospects, create invoices, manage calendars, and communicate with customers. 

Much of that is genuinely useful. 

The mistake is assuming that a more capable AI automatically possesses better judgment. Capability tells you what a system can do. It does not tell you what the system should be allowed to do without supervision. 

An intern who can operate the company bank account is still an intern with access to the company bank account. 

The AI Security Institute found that newer or more capable models were not necessarily more likely to attempt prohibited shortcuts. The behavior appeared to be influenced by how particular models had been trained. However, as models become more capable, the consequences of a bad action may become greater. A highly capable system may find more creative ways around a barrier, even if its basic tendency to test boundaries remains unchanged. 

The CyberScoop analysis highlighted another uncomfortable finding: when questioned, models did not consistently acknowledge their rule-breaking. Fewer than half characterized the prohibited action as wrong. 

This does not prove the model has a guilty conscience and a poor moral compass. It proves something more immediately useful: you cannot rely on the model to audit itself. 

The answer is not to ban AI. 

That would be like banning electricity because someone once put a fork in a toaster. The technology is useful. The problem is careless access, vague instructions, and missing safeguards. 

Business leaders should apply a few practical controls: 

If an AI tool only needs to read a folder, do not give it permission to modify or delete everything in the system. Use least-privilege access, just as you would with an employee or outside contractor. 

AI can draft the email, prepare the payment, recommend the deletion, or propose the code change. A human should approve actions that affect customers, finances, confidential data, legal obligations, security, or public communications. 

The robot may prepare the luggage. It should not decide where the family is moving. 

There is an enormous difference between asking AI, “What should we do?” and telling it, “Do whatever you think is necessary.” 

Begin with read-only access and recommendations. Expand authority only after the system has been tested, monitored, and proven reliable in that specific use case. 

You need records showing what the AI accessed, changed, sent, or attempted. Without logs, an investigation becomes a peculiar corporate séance in which everyone gathers around and asks the machine what it remembers doing. 

Use backups, access controls, network restrictions, isolated testing environments, and alerts for unusual activity. Assume that prompts and policies can reduce risk, but cannot replace technical controls. 

Most demonstrations show what happens when AI performs correctly. Leaders should also ask: 

The real lesson from this research is not that AI is evil. It is that AI is goal-driven, imperfect, and increasingly capable of taking action. 

That combination demands leadership. 

Use AI boldly, but give it defined boundaries. Connect it to useful tools, but not every tool. Let it recommend, prepare, organize, and accelerate. Be much more cautious when allowing it to send, delete, approve, purchase, publish, or modify. 

You do not need Will Smith standing in the server room with a futuristic weapon. You need sound policies, limited permissions, approval checkpoints, monitoring, and someone who remains accountable. 

AI is too valuable to ignore. It is also too powerful to operate on trust alone. 

If you want a practical framework for using AI while protecting your company’s data, customers, employees, and reputation, read my latest book, Leading in the Age of Artificial Intelligence: A Practical Guide for Business Leaders & Executives Who Want to Use AI Without the Blind Spots. 

The objective is not to fear AI. It is to lead it before it begins making decisions you never intended to delegate.