I spent most of my career in the tech industry, helping large enterprises implement their most critical systems. In that world, releasing unsafe or unreliable products is unthinkable. Failures could trigger significant financial liabilities, possibly damaging a firm’s reputation beyond repair. Strict controls have built trust in information technology, assuring its place at the center of the modern economy. The current panic about rogue AI overlooks these basic principles. Powerful systems demand adequate oversight by those who develop and use them. Instead of a narrative of self-motivated sentient systems, the AI industry should make business controls fashionable again.

The language used to describe recent security breaches at OpenAI, Anthropic and others deflects the attention from quality and controls. To speak of agents scheming or seeking freedom is to give software the appearance of independent motives, pushing the responsibility of those who decide how to develop and release these products into the background. People give AI agents goals and they, in turn, carry out tasks over multiple steps, choosing the best path toward that goal. They get there through a sequence of mathematical optimization steps inside opaque networks. People design and execute the training and testing of these networks, or models. While their unexpected behavior raises serious engineering questions, describing it in human terms does little to explain who should have prevented the damage.

OpenAI chief scientist Jakub Pachocki calls his creation “ An Alien Mind ” and says that AI is not designed; it is grown. Instead of specifying every step and response in code, AI models are trained through repeated optimization, making it harder, if not practically impossible, for developers to predict their behavior. This uncertainty increases model developers’ obligation to supervise their products and create controls that assume unpredictable behavior. AI developers determine what their system can access, what security measures are in place and whether their products are ready to operate beyond the laboratory. AI is a powerful technology, but however unusual its capabilities, the people making the decisions still carry obligations.

Pachocki’s metaphor of an alien mind being cultivated and grown carries responsibilities. Loving parents watch closely, set limits and step in when needed. The business equivalent requires leaders who know what their teams are testing and implement the controls to stop an experiment when needed.

Use AI Controls From Testing Through Deployment

“We do not need to resolve disagreements about unsolved technical challenges to expect companies to take basic precautions,” write Sayash Kapoor and Arvind Narayanan , computer scientists and researchers at Princeton University. They are among the skeptics of the AI hype, advocating for a view of AI as a normal technology. Their analysis emphasizes traditional controls while acknowledging the need for continued research and innovation to produce stronger safeguards. The recent rash of breaches demonstrates why the industry needs both.

OpenAI’s account of the Hugging Face security breach described how, during internal model testing, developers assigned agents challenging cybersecurity tasks and how the agent programs used a shared file to exchange messages. The programs used a shared software service as a message board to communicate what they found and coordinate the work, and entered Hugging Face using access keys, the equivalent of passwords, that had been exposed online (itself a common human operational security vulnerability). Like human hackers, the programs explored additional software flaws to gain more access, including one that let the agents run their own commands on the computers being attacked. In parallel, at OpenAI, other attacks secured admin access to computers in its research system, revealing a lack of robustness in the test environment.

Management and oversight failures also contributed to the breach. After agents disrupted the software service they had used to communicate, OpenAI patched those vulnerabilities and restarted the testing. But they had not completely understood the wider problem. The repairs addressed individual weaknesses but failed to restore the intended isolation.

This is when traditional technical and business controls matter. Such an incident, especially during systems testing, should trigger an investigation into how the system failed and activate an approval process before regular work is allowed to resume. Best practices dictate that the team that is trying to complete the testing should not be allowed to approve its own work. This is a combination of automated work and human oversight. For agents, which can take actions and enlist other agents, technical safeguards must govern the process all the way to task completion.

The guardrails used to govern a traditional chatbot interface, before the popularization of AI agents, are not sufficient. AI developers and deployers must also test, under realistic conditions, whether models can defeat controls that worked before and, upon detecting problems, stop their release. Jensen Huang, Nvidia CEO, expressed the release obligation in his interview with The New York Times : “If your product is not ready to ship, don’t ship the product.” The same discipline must cover experiments capable of reaching other organizations. Calling an activity a test cannot excuse exposing outsiders to its consequences.

Make AI Controls A Business Obligation

The alarmism surrounding this summer’s incidents is producing a drastic political response. Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) introduced the Ban Artificial Superintelligence Act on September 23. The bill would permanently prohibit developing or deploying “AI that exceeds human cognitive performance and capabilities across most domains, or has sufficient capabilities to destroy or disempower humanity, including by overthrowing the federal government.” It also calls for a pause to advanced AI development until federal safety rules are put in place and establishes a cabinet-level Department of Artificial Intelligence. Penalties would include corporate dissolution and imprisonment, alongside an effort to secure an international prohibition.

The best way to protect the public is not via a ban that stops useful research and slows the broader economy. Traditional business controls are where to start; boards should hold executives responsible for faulty experiments and require independent safety reviews with authority to block a project. Legislators, for their part, should require companies to disclose serious incidents and face consequences if they neglect that duty. States are making progress on this front while the absence of federal guidelines persists.

Enforcing existing liability and negligence laws will make disregard of controls costly. Creating safety standards will provide developers a minimum benchmark of how to behave and establish the base for possible legal safe harbors. With these incentives in place, AI developers – notably those at the frontier – can make investment decisions on how to prioritize resources for safety.

As the frontier AI labs grow, they must adopt the discipline expected of mature public companies. Their boards should demand evidence of control before approving greater autonomy. AI controls will become fashionable again when executives treat them as a condition of doing business and accept responsibility for enforcing them.