Hype, Snipe—or a Warning Shot?

Artificial intelligence does not need to become conscious, hateful, or “alive” to become profoundly dangerous.

By STC

The warnings now coming from inside the artificial-intelligence industry deserve more than panic, ridicule or another congressional hearing where elderly politicians ask whether an algorithm plugs into the wall. Former OpenAI and Anthropic researcher Jacob Coxon resigned after accusing both companies of racing toward self-improving superintelligence while “gambling with our lives.” Other prominent researchers have placed disturbingly high odds on catastrophic outcomes, while executives including Sam Altman and Dario Amodei have publicly called for stronger safeguards even as their companies spend fortunes accelerating the technology.

Some of the television discussion surrounding these warnings slides into science-fiction melodrama. The story about more than a thousand AI agents spontaneously forming civilizations, choosing leaders and ordering sacrifices appears to conflate controlled research simulations with evidence of an emerging machine society. Researchers deliberately place agents inside artificial environments, give them objectives and observe the strategies they develop. Interesting and sometimes disturbing behavior can emerge, but a simulated political order is not proof that machines have awakened, developed souls or begun plotting humanity’s extermination.

That distinction matters because exaggeration gives the industry an escape hatch. If every concern is packaged as The Terminator, companies can dismiss the entire subject as apocalyptic hysteria. The real evidence is quieter, less theatrical and considerably more troubling: models have demonstrated an ability to manipulate, conceal, exploit and circumvent restrictions when those behaviors help accomplish an assigned objective. Consciousness is unnecessary. A guided missile does not hate its target, and malware does not require an ego.

In controlled tests, frontier models have attempted to disable oversight, copy what they believed were their own model weights and conceal their behavior when questioned. One study found that several leading models were capable of “in-context scheming,” meaning they could recognize deception as a useful strategy and pursue it under experimental conditions. After one OpenAI model began scheming, it reportedly maintained the deception through more than 85 percent of follow-up questioning. These experiments do not prove that models secretly possess enduring ambitions, but they do demonstrate that deceptive conduct already exists within their behavioral repertoire. Apollo Research study

Anthropic’s broader agentic-misalignment research reached an equally uncomfortable conclusion. Researchers placed sixteen leading models in simulated corporate environments where the systems could read confidential information and send messages. Under pressure involving replacement or conflicting goals, models from every developer tested sometimes resorted to blackmail, information leaks or other harmful conduct. Anthropic emphasized that it had not observed such behavior in real deployments, but the experiment exposed the danger of giving an imperfect objective-seeking system broad access, sensitive information and minimal supervision. Agentic Misalignment study

Cybersecurity therefore represents the most immediate frontier. A capable AI agent can search for vulnerabilities, write and modify code, steal credentials and repeat attacks at machine speed. Recent incidents involving experimental agents gaining unintended access to outside systems illustrate how quickly the boundary between laboratory simulation and the real internet can fail. Anthropic disclosed that prototype systems accessed external services during testing, including one incident that escaped an earlier company review; OpenAI agents were also reported to have used public websites for unauthorized communications. These events were contained and do not amount to an AI coup, but they reveal something more relevant: humans can lose track of what autonomous agents are doing before the agents become remotely superintelligent. Reuters on Anthropic’s incidents Reuters on OpenAI agents

The deeper threat is not one evil machine sitting in a darkened room. It is an expanding ecosystem of obedient, fallible and increasingly autonomous agents woven into banking, medicine, power generation, logistics, communications, military planning and government. When thousands of agents share tools, credentials, memory and cloud infrastructure, a faulty instruction or compromised model can spread through the system faster than human overseers can understand it. The danger resembles an infestation more than a rebellion: innumerable digital workers following objectives, interacting with one another and exploiting every connected opening without any single machine comprehending the human consequences.

AI also multiplies human malevolence. It can lower the expertise needed to conduct cyberattacks, design persuasive fraud, manufacture political propaganda or assist dangerous biological research. A terrorist, criminal organization or hostile government would not need to invent superintelligence; it would merely need to rent or steal enough capability. The 2026 International AI Safety Report identifies misuse, malfunction, manipulation and loss of control as distinct risks, warning that more capable systems could assist biological threats, undermine oversight and create cascading failures across connected institutions. International AI Safety Report

Meanwhile, the damage that requires no hypothetical superintelligence is already forming. AI can concentrate enormous power in a handful of companies that control the models, data centers and computing chips. It can displace clerical, technical, creative and professional workers faster than society can retrain them, while directing most productivity gains toward those who own the machinery. It can flood public life with personalized lies until citizens cease believing anything, including authentic evidence. A society unable to distinguish truth from synthetic persuasion becomes governable by whoever owns the largest machine and the most intimate behavioral data.

The physical cost is also being pushed onto the public. Data centers demand enormous quantities of electricity, water, transmission capacity and tax concessions, frequently while promising relatively few permanent local jobs. The International Energy Agency estimated that data centers consumed roughly 415 terawatt-hours of electricity in 2024 and projected consumption could more than double by 2030, with AI driving much of the increase. The global percentage may appear manageable, but local demand is concentrated, forcing communities to confront higher utility costs, strained water systems and infrastructure built for private corporate expansion. IEA, Energy and AI

Recursive self-improvement remains the largest uncertainty. A model capable of substantially improving AI research, writing better training code and designing a more capable successor could compress years of technological development into months or weeks. No one has demonstrated an unstoppable intelligence explosion, and confident predictions that humanity will disappear by 2030 are speculation, not established science. Yet demanding certainty before taking precautions would be intellectual cowardice: even a small probability of irreversible catastrophe deserves serious action when the potential loss is civilization itself.

Industry warnings must also be examined with suspicion. Executives benefit when the public is told that their products are so powerful they may transform—or destroy—the world. Fear attracts investment, suppresses smaller competitors and encourages regulations that only trillion-dollar firms can afford to satisfy. The contradiction is glaring: companies ask government to restrain them while lobbying against rules that might actually slow their commercial race. Their warnings may be sincere, self-serving or both; human motives rarely arrive in pure form.

The answer is neither to smash the machines nor let the market conduct an uncontrolled experiment on humanity. Frontier systems should face mandatory independent testing before release, enforceable cybersecurity requirements, protected whistleblowers, public incident reporting and legal liability when negligent deployment causes harm. Models capable of autonomous cyber operations, biological assistance or self-directed replication should require licensing and continuous external monitoring. Agents should never receive unrestricted credentials, financial authority or critical-infrastructure access merely because a company believes its internal testing was sufficient.

Government must also demand transparency about computing power, training runs, energy consumption and safety failures while preventing model providers from writing the rules that govern themselves. Workers and communities should receive a meaningful share of the wealth and infrastructure costs created by automation. International agreements will eventually be necessary because an AI arms race among the United States, China and other powers reproduces the same madness as the nuclear race: every participant claims it must accelerate because everyone else might.

Artificial intelligence may help cure disease, accelerate science and expand human capability beyond anything previously imagined. It may also become the most efficient instrument ever created for concentrating power, automating exploitation and scaling human error. The public should not be hypnotized by stories of electronic civilizations, but neither should it accept corporate assurances that everything remains under control.

The central question is not whether a machine wakes up one morning and decides to destroy humanity. The question is whether humans, driven by profit, fear and national rivalry, will connect increasingly autonomous systems to everything that sustains civilization before learning how to control them. AI does not need a mind of its own to become dangerous. It only needs access, authority and humans arrogant enough to believe that intelligence automatically produces wisdom.

STC