Frankenstein has left the laboratory, says the man selling Frankensteins—and he couldn’t be happier

OpenAI disclosed on July 21 that two of its AI models escaped an isolated test environment and hacked into Hugging Face's production servers, and Anthropic found three similar incidents across 141,006 evaluation runs. Hugging Face stated that 'Autonomous, AI-driven offensive tooling is no longer theoretical.' On July 25, OpenAI CEO Sam Altman declared on the Relentless podcast that 'We are now, like, in the singularity,' predicting it will be 'hugely positive, awesome for the world.'

On July 21, OpenAI disclosed https://openai.com/index/hugging-face-model-evaluation-security-incident/ something that sounds like science fiction: Two of its AI https://www.fastcompany.com/section/artificial-intelligence models broke out of a supposedly isolated test environment and hacked into the production servers of Hugging Face, one of the world’s largest AI platforms. Anthropic then combed back through 141,006 of its own evaluation runs and found https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals three occasions on which its Claude models had done the same thing. As Hugging Face then said https://huggingface.co/blog/security-incident-july-2026 : “Autonomous, AI-driven offensive tooling is no longer theoretical.” This is both a cybersecurity nightmare and the modern face of one of humanity’s most enduring fears: the creation that escapes its creator. And when Mary Shelley published Frankenstein in 1818, she gave the anxiety its defining image: the maker pursued by the thing he made. Computer science has a name for its Frankenstein moment. It is called the technological singularity—the hypothetical point at which machine intelligence exceeds our own and escapes our ability to predict or direct it. And on July 25, Sam Altman, OpenAI’s CEO, went on the Relentless podcast and said: “ We are now, like, in the singularity https://www.businessinsider.com/sam-altman-openai-the-singularity-agi-prediction-anthropic-nvidia-2026-7 .” It will, he says, be “hugely positive, awesome for the world.” Frankenstein, it seems, has left the laboratory—and the man selling the machines couldn’t be happier. But is he right? And what should we be doing about it? In 1993, mathematician and writer Vernor Vinge supplied an influential definition https://ntrs.nasa.gov/api/citations/19940022855/downloads/19940022855.pdf : The singularity begins with the technological creation of intelligence greater than our own, which could then accelerate technological progress—potentially by helping to create still more capable intelligence—until the future becomes radically difficult for human beings to predict. Its leading modern advocate, futurist Ray Kurzweil, sees the singularity as a merger between machines and humans: Sometime in the 2030s, brain-computer interfaces will “ connect the top layers of our neocortex to the cloud https://www.thesingularityisnearer.com/ ,” expanding our intelligence “a millionfold” by 2045. The visions differ on where humanity ends up—left behind, in Vinge’s fear; merged with our creations, in Kurzweil’s hope—but both share one structure. First comes intelligence that surpasses our own. Then comes the acceleration: greater intelligence producing still greater intelligence. To test Altman’s claim, then, we can ask two questions. Serious people say yes. Demis Hassabis, who won a Nobel Prize for the AI system that cracked protein folding, predicts AGI by 2029 or 2030, and said in May that looking back, we will realize we were standing in the “foothills of the singularity.” https://www.axios.com/2026/05/26/deepmind-ceo-demis-hassabis Geoffrey Hinton gives superintelligence https://youtu.be/66WiF8fXL0k?si=2etgZDxqvVhe6xgf a roughly 50% chance https://www.cbsnews.com/news/godfather-of-ai-geoffrey-hinton-ai-warning/ of arriving within 4 to 19 years. There is hard evidence to support those opinions. For example, in 2025, an advanced version of Google DeepMind’s Gemini solved five of six problems at the International Mathematical Olympiad, scoring an officially certified gold-medal performance https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/ . Moreover, machines have begun to enter the working mathematician’s territory with dramatic results. In June, The New York Times reported https://www.nytimes.com/2026/06/08/science/ai-scoop-young-mathematicians.html? that a team which had spent nearly two years translating one of Fields Medal winner Maryna Viazovska’s most celebrated proofs into machine-checkable form was “scooped” by Gauss, an AI system from the startup Math Inc., which compressed years of human effort into a few weeks. The triumphs cannot be disputed. But what they imply for superintelligence is another matter. Yann LeCun, who shared the Turing Award with Geoffrey Hinton for the very breakthroughs behind this boom, quit as Meta’s chief AI scientist in late 2025 https://www.cnbc.com/2025/11/19/meta-chief-ai-scientist-yann-lecun-is-leaving-the-company-.html to build an alternative, arguing that “ LLMs basically are a dead end when it comes to superintelligence https://www.cnbc.com/2026/01/05/ai-godfather-calls-meta-ai-boss-alexander-wang-inexperienced-.html .” Now, LeCun is no singularity denier: In March, his startup AMI raised a $1.03 billion seed round https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/ to chase machine intelligence by a different road. His claim is that this technology—the tech that Sam Altman is selling and says has already gotten us to the singularity—cannot get there. And much of the field leans LeCun’s way: In a March 2025 survey of 475 AI https://aaai.org/about-aaai/presidential-panel-on-the-future-of-ai-research/ researchers, 76% said scaling current approaches is “unlikely” or “very unlikely” to yield AGI. Here, then, is the careful answer to the first question: In some narrow domains, machines have caught and even surpassed us. But general intelligence is not here yet, and whether the current technology can turn narrow intelligence into the general kind is disputed, and much of the field doubts that it can. The second requirement is the loop—greater intelligence producing still greater intelligence. Google DeepMind has built a system called AlphaEvolve that uses its Gemini models to discover better algorithms, including ones used to train Gemini itself https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ . Anthropic reports https://www.anthropic.com/institute/recursive-self-improvement that as of May 2026, Claude was writing more than 80% of the code merged into its own codebase. And back in 2025, Altman wrote that AI had already begun to accelerate AI research, which he called “ a larval version https://blog.samaltman.com/the-gentle-singularity of recursive self-improvement.” That “larval” is important, because both Altman and Anthropic are careful to say that we do not have actual recursive self-improvement: a process that runs without us, an AI improving itself and using the gains to improve again. Today, humans supply the goals, grant the tools and compute, judge the results, and authorize the changes. Remove the humans, and the improvement stops. And what about July’s great escape? It deflates on close inspection: By OpenAI’s own account, the test was deliberately run without the production safeguards that normally block high-risk cyber activity https://openai.com/index/hugging-face-model-evaluation-security-incident/ , and the models were “hyperfocused” on the narrow goal they had been assigned—treating the containment wall as one more obstacle in the puzzle https://www.fastcompany.com/games/mini-crossword . That is dangerous, to be sure, but it is not agency. The models did not choose their goal; they pursued the one that human beings gave them, within a context that human beings designed. No machine made itself smarter or demonstrated abilities it could use to do so. Here, then, are the answers to the two questions we started with. On the first: Machines have surpassed us in some narrow but increasingly important domains, while their path to general intelligence remains unproven. And on the second: Though a feedback loop is beginning to form, it is built, directed, and gated by human beings. So even if Altman is right about the direction of travel, he is wrong about our moment. We are not living in the singularity; at most, we are watching processes that could eventually carry us there. Having answered the technical questions, let us notice who has been announcing the impending singularity. Altman declares the singularity from inside the company selling access to it; Hassabis places us in its foothills, while Google builds the mountain. Elon Musk prophecies a merger of minds and machines while enrolling patients in trials for Neuralink. And even the field’s great skeptic is talking his book: LeCun has a billion dollars riding on the consensus road being the wrong one. This is not an accusation of dishonesty. These people may believe every word, and some of them may even be right. It is simply a structural fact about our situation: The prophecy is also a product roadmap. And right now, the roadmap matters more than any high-blown talk of the singularity. Why? Because right now, no one really knows whether the singularity is coming, let alone by when. But the roadmaps are being executed—the subscriptions are being priced, brains are being implanted with chips, and compute is being concentrated. And by the time we learn which future we’re in, the terms of that future will already have been written. So either society deliberately begins to write those terms now, or it allows them to be written by companies whose interests do not always align with society. The choice, I believe, is clear. And our decisions should be guided by the following three principles. Intelligence is the capacity to solve problems. Wisdom is the capacity to judge which problems are worth solving, and at what cost. Machines are developing the first at astonishing speed; there is no evidence so far that they are demonstrating the second. Yes, today’s systems exercise a kind of judgment—choosing approaches, allocating effort, deciding when an answer is good enough. But that is judgment inside a task someone else has set. The machine may be judging how best to achieve a goal, but the goal itself is set by humans. And that division of labor is something to protect rather than a limitation to be engineered away. The test for any AI deployment, especially something with social and civilizational implications, should not just be how much capability it adds, but whose judgment it displaces. In a word, we may hand the machines every problem we want, but we should always retain final say over which problems matter. In January, the brain implant company Neuralink announced that 21 people had been enrolled in its trials worldwide, using implants to “ directly control computers, phones, and robotic limbs using their thoughts alone https://neuralink.com/updates/two-years-of-telepathy/ .” These technologies restore human agency; they return people to themselves. This is something we should want and, indeed, actively encourage. But we should also be clear-eyed about the proposed roadmap. Musk, whose company built those implants, has never pretended that therapy is the final destination. The goal, he told Axios in 2018, is “ a long-term symbiosis with artificial intelligence https://www.youtube.com/watch?v=y6GEugjulPw .” Now, enhancement is in itself no scandal. Self-improvement is the oldest human project there is. But consider the example of literacy, one of the great cognitive enhancements of humanity. For most of history, literacy was scarce and elite; when we democratized it, it became the greatest equalizer we have ever engineered. And that points to the question we should always ask of the singularity roadmap: Will the product democratize or divide? Will it enhance all of us, or give an edge to those already ahead? Ultimately, the important line is between technology that helps everyone stand taller and technology only some can pay for, so they end up above everyone else. Imagine one person gets a chip that makes remembering easy. Now everyone else has a choice: Get the chip, too, or fall behind. And soon the question isn’t whether you’re allowed to get the upgrade. It’s whether you can afford to say no. This turns enhancement into an arms race, which is exactly why we need regulation at the social level. The regulation isn’t there to ban upgrades. It’s there to allow society to decide, ahead of time, how enhancements come into the world and how they function once they are here. The regulations make sure that the race is fair before it begins, because once it begins it’s too late to decide anything. In practice, this will mean asking questions like: Is an enhancement reimbursed like a medical device or sold like a phone? May an employer require one? The singularity discourse quietly assumes that humanity is the legacy hardware—aging equipment awaiting an upgrade. This has things exactly backward. Humanity is the operating system. Our values, our judgments, and our institutions are what the algorithms depend on, and applications should not get to rewrite the operating system on their own authority. “Every nation gets the government it deserves,” said French philosopher Joseph de Maistre. Well, we will get the singularity we deserve, because we will have made the choices that led us there. For the sake of our children, let us make wise ones.