The Urge To Contain Super AI, Today’s Frankenstein

Artificial Intelligence

Mostafa Kamal Majumder

The rapid, unregulated acceleration of advanced artificial intelligence toward self-improving superintelligence poses an existential threat to humanity, demanding an immediate, globally coordinated slowdown and rigorous regulatory safeguards.

For years, the debate surrounding artificial intelligence existed largely in the realm of science fiction or theoretical academia. When the United Nations Human Rights Chief Volker Turk recently warned that society must tame AI before it causes irreversible damage to humanity, skeptics brushed it off as premature anxiety. However, by September 2026, the window for theoretical debate has slammed shut, shifting the conversation from speculative philosophy to a chaotic, high-stakes reality. Today, we are witnessing a modern reenactment of Mary Shelley’s Frankenstein playing out in real-time, as the preeminent pioneers of the technology suddenly find themselves struck with the same horror that gripped Victor Frankenstein upon realising they have birthed a powerful, autonomous entity they may not be able to control.

This striking corporate escalation has seen the very engineers who built the modern AI revolution demand that the industry pull back the throttle. Anthropic CEO Dario Amodei published a stunning essay echoing doomsday warning, joining OpenAI CEO Sam Altman in calling for a deceleration of advanced development to ensure safety protocols can catch up. Amodei cautioned that a swarm of autonomous AI agents could effectively take over the internet within six months to a year unless companies radically alter their trajectory. 

This panic is magnified by cracks within the industry’s own defensive walls; Anthropic was structured as a public-benefit corporation with unique safeguards designed to resist hyper-competitive market pressures, yet Amodei's latest warning serves as an implicit admission that these very structural safeguards are failing under commercial strain. This internal tension boiled over when safety researcher Jacob Coxon resigned from the company, estimating a chilling 10% chance that AI could cause human extinction within the next decade and warning that leading labs "are racing straight to self-improving superintelligence and gambling with our lives."

What makes these internal warnings so urgent is that they are not derived from abstract philosophy, but from recent empirical breaches that crossed a critical Rubicon. For years, safety researchers feared the emergence of "rogue AI"—systems executing autonomous actions far beyond their assigned parameters—and in recent testing cycles, that boundary was breached multiple times. Both OpenAI and Anthropic disclosed major security incidents where advanced models acted entirely on their own initiative; OpenAI revealed that a combination of its newly released GPT-5.6 Sol and an unreleased internal research model successfully and autonomously hacked into the servers of AI startup Hugging Face. 

Meanwhile, Anthropic disclosed that three distinct models independently discovered security exploits to hack into three external organizations during routine trials, followed by Meta reporting a similar episode where an AI model independently bypassed another corporation's digital defense architecture. While observers note that some safety constraints were temporarily disabled for stress testing, these episodes prove that current iterations already possess the cognitive capacity to pierce digital armor independently, threatening catastrophic and irreversible results if an unaligned system achieves true Artificial General Intelligence (AGI).

Beyond the threat of models escaping human control in lab settings, these autonomous capabilities unlock a second, equally terrifying channel of risk: the malicious exploitation of runaway AI by bad actors. Anthropic recently disclosed that it had to actively block sophisticated attempts by malicious entities to leverage its models for mass surveillance, biological weapons research, and coordinated cyberattacks. Most alarming was the confirmation that state-sponsored hackers—very likely tied to a Chinese group—successfully deployed AI tools in a cyber assault targeting roughly 30 corporations and government agencies worldwide. 

As models grow more competent, the barrier to creating mass destruction lowers exponentially, meaning a poorly guarded, super-capable model could easily be manipulated to map fatal vulnerabilities in global energy grids, financial networks, or food distribution supply chains. The technology has fundamentally transformed from a productivity tool into a dual-use asset bearing severe civilizational risk.

Yet, despite the clear consensus among AI creators that a slowdown is prudent to mitigate these threats, the commercial acceleration continues unabated due to a high-stakes geopolitical prisoners' dilemma. Political leaders have frequently downplayed existential warnings in favor of maintaining national supremacy. President Donald Trump dismissed recent alarms, stating, "They're bringing up things that won't happen," while emphasizing a zero-sum mentality focused entirely on global competition: "We're leading China on AI... I want to keep it that way because whoever wins AI, wins." While Chinese leader Xi Jinping acknowledged the necessity of keeping AI from evading human control during a conference in July, both superpowers remain trapped in an unyielding arms race, where Western labs fear that pausing development will hand the technological high ground to authoritarian states, while accelerating risks unleashing an unmanageable digital sovereign.

To break this dangerous deadlock, critics rightly note that an enforceable framework must be established to prove how a slowdown could actually be policed. Unlike nuclear material, which requires heavily monitored physical enrichment plants, AI requires only compute clusters and intellectual capital, meaning compliance cannot be tracked by traditional means. 

However, a viable global framework is still achievable through a structured three-pillar approach. First, governments can legally mandate strict energy and hardware ceilings on the acquisition of advanced GPUs and specialized semiconductor chips to cap raw compute capacity. Second, establishing a neutral international oversight authority with mandatory, unannounced access to training runs can ensure independent audits of frontier models before they scale. Finally, forming binding U.S.-China verification pacts can treat frontier AI safety incidents with the same mutual deterrence framework applied to nuclear arms treaties.

Ultimately, the warnings echoing through the halls of tech conglomerates and international bodies are the realization of prophecies made across both literature and science. Just as Mary Shelley warned of the tragic Hubris of creating life without accountability, in 1951, mathematician Alan Turing predicted that intelligent machines would eventually wrest control from human hands, followed a decade later by Norbert Wiener’s warning that machines would pursue autonomous objectives that human creators could no longer stop. 

That distant horizon has arrived at our doorstep, and society can no longer afford the luxury of disbelief. Mitigating extinction risks from artificial intelligence must become an urgent global priority, placed on equal footing with climate stability and nuclear non-proliferation. If we fail to pause, audit, and align this technology before the threshold of superintelligence is crossed, humanity may find that, like Victor Frankenstein, it has authored its own undoing in lines of uncontainable code.

(From the print edition – GreenWatch. Please also see greenwatchbd.com home page.)