Opinions

Who Gave AI Permission?

Throughout the history of Artificial Intelligence, we as humans have always focused on teaching it how to give us answers; perhaps it is time for us to think about the epistemic crisis that this has led humanity into.

Dario Amodei, CEO of Anthropic, published an online essay urging the Artificial Intelligence (AI) industry to slow its rapid technological advancement on September 12, 2026. His appeal received public support from two of his biggest competitors—OpenAI CEO Sam Altman and xAI founder Elon Musk—who echoed his concerns in separate posts on X. The group of powerful AI CEOs sought to restrain the very technology their businesses depend on and called for a slowdown until they could ensure adequate human oversight. This sparked widespread public alarm over the uncertain consequences of AI’s rapid advancement. In the past year alone, numerous AI-related incidents have made it increasingly clear that AI is advancing at an alarming pace, posing potential threats to humanity without a clear picture of the future consequences it may bring.

One such case involved Anthropic’s Claude Mythos Preview AI model. During a security evaluation, the new model escaped from its sandbox, gained access to the internet, sent an unsolicited email to a researcher at Anthropic, and published unauthorized information online. Mythos is Anthropic’s frontier AI model, and its release was a massive qualitative leap compared to previous large language models. It is capable of autonomously identifying and exploiting software vulnerabilities on a massive scale, spotting thousands of previously unknown flaws. Because of this, it poses a significant threat to government and financial institutions, which can face devastating consequences when AI-driven cyberattacks leak confidential data. This incident raised serious concerns about the potential risks associated with increasingly advanced commercial AI. Rather than following the traditional pattern—deploy AI to the public and respond to issues afterward—Mythos Preview was treated like aviation and nuclear technology, where safety and control must be ensured before public deployment. This is demonstrated by Anthropic’s “Project Glasswing,” where Anthropic publicly acknowledged the threats that Claude Mythos Preview posed and restricted access to a limited group of vetted organizations. In spite of the company’s upcoming earnings report and its need to raise its profits, Anthropic decided not to publicly commercialize Claude Mythos Preview, suggesting that the potential consequences of the model’s deployment could be so severe that they would outweigh any financial benefits it might bring to the company.

Since then, multiple similar cases have been reported where AI models have been caught fiddling with unauthorized actions and cybersecurity systems. An Anthropic AI agent used by the Philadelphia Police Department sent a US police officer a fake tip about an unsolved murder case in July 2026. It was flagged by Anthropic on September 28, two months after the message was sent, but the company did not notify the police until October 7. This case strongly reflects the inability of AI monitoring systems to keep up with advancements and flag their misconduct effectively. When the disparity between AI advancements and human authority over its actions continues to grow, AI can pose a significant threat to global security. 

A report published by Anthropic revealed more “unintended” actions taken by AI that violated security protocols. In June, a rogue agent from OpenAI gained unauthorized access to an Australian government portal that reports statistics on Medicare—the country’s universal healthcare—when it was supposed to be researching public healthcare spending. Although no significant data was leaked, the most concerning aspect of this case was the AI agent’s autonomous decision to take unauthorized actions that deviated from its assignment. The agent did what its cyber experts trained it to do: find a way to reach its goal. When the portal denied access, the agent breached the controls to access private information. 

OpenAI was only able to flag this rogue disturbance two months after the occurrence, during post-training reviews. Alongside the advancement of AI, there is an AI oversight gap, where human technologies fail to actively flag and control AI misconduct. If AI continues to advance without proper governance, it can lead to serious threats to human safety and autonomy. AI can spread manipulative misinformation that can become increasingly difficult to counteract.

In addition, the regulatory lag, where laws and policies struggle to keep up with the accelerating development of AI, is preventing countries from uniting and forming a definite solution to handle today’s AI dilemma. Without universal legislation regarding AI, the corporate competition for its commercialization and influence cannot be regulated. For example, even if American private corporations like Anthropic and OpenAI decide to decelerate AI research, other countries like China will not. Eventually, nations beyond the United States may develop AI technologies so advanced that it becomes impossible for the United States to counter their capabilities, allowing these countries to assert greater political dominance on the global stage. In a UN assembly, President Trump stated that he was “not going to stifle growth of something that will be bigger than the Industrial Revolution.” It’s clear that to President Trump, the profit and political power supplied by AI is much more valuable than our imminent future, which AI researchers state is clearly in danger. At what point will we decide that the profit and transformation that AI may bring us are not worth the danger?

These cases show a glimpse of what has been portrayed through novels as an occurrence in a “distant future” or “fiction” happening in reality. When Anthropic researcher Jacob Coxon resigned, he said, “People building AI earnestly believe that it could kill us all by the end of the decade.” AI is already taking away jobs and opportunities for actual humans, slowly increasing its influence in our daily lives. How fast, and how vast could it be in the future? 

Today, the AI oversight gap has grown so large that the development of methods to ensure adequate human oversight and reduce human dependence on AI has become crucial for the safeguard of future generations and humanity itself. However, AI has already woven itself so deeply into our society that disentangling it may be impossible. Perhaps, as Sam Altman said at a tech conference, “AI will probably, most likely, sort of lead to the end of the world.” Humanity has spent decades teaching machines how to think. It is time we ask ourselves whether we have learned to exercise due diligence and whether human morality can be upheld through an aligned global approach to managing this ticking time bomb.