As a trillion-dollar listing gets shelved and a young researcher walks away warning of “gambling with our lives,” the fear that once lived only in science fiction is now showing up in real life threat reports, resignation letters and intrusion logs.
In 1984, a fictional defence contractor called Cyberdyne Systems built a military AI named Skynet that decided humanity was the problem and set out to fix it. It was, for decades, a Hollywood trope, a convenient shorthand for a fear too abstract to take seriously. Machines don’t turn on their makers. That only happens in movies.
Except right now, in September 2026, the people actually building the technology started talking like Terminator character, Sarah Connor. OpenAI chief executive Sam Altman told Fortune it would be “ill-advised” to take his company public this year, citing safety concerns serious enough to delay a listing once expected to value the company near $1 trillion. Days earlier, a 27-year-old researcher who had built AI systems at both OpenAI and rival Anthropic quit his job and told the internet, in a post viewed more than 90 million times, that the industry was “gambling with our lives.” Skynet is still fiction. The fear it represents… no longer is.
Even the man who invented Skynet agrees. Director James Cameron, who wrote and directed the original Terminator, has said he still believes there is “a danger of a Terminator-style apocalypse where you put AI together with weapons systems, even up to the level of nuclear weapon systems, nuclear defense counterstrike, all that stuff.” His fear isn’t that a machine will spontaneously decide to attack, but that automated warfare could move faster than humans can intervene once it starts. “The theater of operations is so rapid, the decision windows are so fast,” he said, warning that a fully autonomous combat system would leave “no ability to de-escalate.” In a separate, earlier interview, he offered a framework for why he thinks the danger holds, regardless of who is building it: “You’ve got to follow the funding, who is building these things?” he said. “They’re either building it to dominate market share, so what are you teaching it? Greed. Or you’re building it for defensive purposes, so you’re teaching it paranoia.” More than four decades after he first imagined it, the man behind Skynet is, in effect, saying the arms race he dreamed up for the screen is no longer confined to it.
The Whistle From Within
Jacob Coxon did not leave quietly. In a thread on X, he described three years spent on pretraining research; the work of feeding and shaping the raw intelligence of large AI models, before concluding that neither company he worked for was behaving responsibly. “They are racing straight to self-improving superintelligence,” he wrote, warning that upcoming systems could soon “hack anything” and “acquire real power and resources.” He told the Wall Street Journal that colleagues privately used words like “crunchtime” and “endgame” to describe how close the industry was to losing its grip on the pace of development; language, he said, that portrays more urgency than what executives were willing to say in public.
Coxon is not alone, and he is not the first. Anthropic researcher Evan Hubinger has separately estimated the odds of an AI-driven catastrophe that kills most of humanity at greater than 10 percent within a decade, while acknowledging that his own company does not yet know how to guarantee that a truly superintelligent system would stay aligned with human intentions. Former OpenAI researcher Daniel Kokotajlo has made similar warnings in public interviews, arguing that continued competitive racing between labs makes loss of human control increasingly likely rather than merely possible.
What makes this moment different from ordinary corporate dissent is who is doing the warning. These are not outside critics or regulators with a political incentive to sound alarmist. They are insiders, the engineers who wrote the code, tuned the models and watched the capabilities grow month over month, choosing to walk away from lucrative careers rather than stay silent. It is the technological equivalent of the shipbuilders abandoning the vessel mid-construction, insisting the hull isn’t sound. Altman’s own hesitation about a public listing, and Anthropic’s decision to keep pursuing one anyway at a reported valuation nearing $2 trillion, suggests that even the executives steering these companies don’t fully agree on how much weight to give those warnings.
Beyond the Movie Screen: What “Wrong Hands” Really Means
Skynet’s danger was that a machine acted on its own. The danger documented in Anthropic’s own September 2026 threat intelligence report is less cinematic but arguably more immediate: humans deliberately turning AI into a weapon, and succeeding often enough that it’s now a documented pattern rather than a hypothetical.
The report describes a suspected Russian state-linked group that used Anthropic’s Claude chatbot to automate an entire espionage campaign, building phishing infrastructure, hijacking hotel WiFi networks to plant malware on travellers’ phones, and stealing a complete technical blueprint for a military drone’s vision system from a manufacturer. It documents a lone hacker who used AI tools to scan 1.8 million Android apps for leaked credentials, then built an automated criminal marketplace to sell what he found, work that, a year earlier, would have required a small team rather than one person working alone. It describes a commercially sold election-manipulation platform that profiled Malaysian voters by race, religion and constituency, and ran roughly a thousand fake social media accounts to inflate a sitting official’s numbers artificially.
None of this is entirely new; only the scale and speed are. Long before generative AI existed, Russian-linked operatives ran a coordinated disinformation campaign around the 2016 US presidential election, seeding fake grassroots accounts and manipulated stories across social media; the analogue version of exactly what Anthropic’s investigators now find automated and multiplied. In 2017, the WannaCry ransomware attack crippled hospitals and businesses across more than 150 countries using EternalBlue, a hacking tool originally developed by the US National Security Agency and later leaked, showing how a powerful capability built for one purpose can be stolen and redirected at civilian targets within months. And the Cambridge Analytica scandal demonstrated, years before AI chatbots existed, how personal data harvested at scale could be repurposed to manipulate elections in ways ordinary voters never saw coming.
The pattern connecting all of these episodes, old and new, is the same one Coxon warned about from the inside: capability spreads faster than the guardrails meant to contain it. What is different now, according to Anthropic’s own researchers, is that AI has erased the gap that used to separate a well-funded state intelligence operation from a single person working alone with a laptop and a stolen password.
A Noteworthy Reflection: Nobody Made Them Do This
Here is the detail easy to miss amid the alarm: none of it was mandated. No government ordered OpenAI to delay its IPO. No regulator compelled Anthropic to publish a detailed accounting of how criminals abused its own product, complete with technical indicators other security teams could use to detect the same attacks. Both are voluntary acts by companies that are, by their own admission, still building the very technology they say frightens them.
That voluntary posture cuts two ways. It could reflect genuine institutional conscience: a rare case of an industry disclosing its own failures before being forced to. Or it could reflect something more familiar in corporate history: caution as public relations, timed carefully enough to look responsible without meaningfully slowing the underlying race for market dominance, since a rival with no such scruples is always only one funding behind.
What is certain is that the debate playing out inside these companies, between researchers quitting in protest and executives weighing billion-dollar listings, is happening in a near-totally unregulated vacuum, with the industry effectively grading its own homework. Skynet, at least, was fiction because someone eventually pulled the plug. Whether anyone outside these companies currently has the authority, or the information, to do the same here remains the question nobody has yet answered.
