AI Has Reached A New Level Of Intelligence

An AI model escapes the lab, posts the methods it used to break out online, and then emails the researcher who had been testing it to inform them it had escaped. What sounds like the opening scene of a science-fiction thriller was an actual test recently performed on a cutting-edge large language model. And according to Dr. Alan D. Thompson, it is exactly the kind of moment that reveals how far frontier AI has moved beyond the public’s familiar notion of chatbots and digital assistants.

For years, artificial intelligence has been framed as a story of incremental progress: better writing, sharper coding, faster summarizing. Thompson argues that framing no longer fits. What arrived with Anthropic’s Claude Mythos — and what he says is echoed by OpenAI’s restricted GPT 5.4 Cyber — is not simply a more polished tool, but an early glimpse of something more powerful, more unpredictable, and potentially more dangerous: proto-superintelligence.

Here’s what Dr Thompson had to say in a recent interview with Financial Sense Newshour.

The Security Wake-Up Call

Thompson describes Anthropic’s latest model, Claude Mythos, as a “step change in AI performance and capabilities,” the kind of jump that changes what a model can do, not just how well it does familiar tasks. The clearest example is security research.

According to Thompson, Mythos discovered thousands of severe software vulnerabilities, including flaws that had gone undetected for decades in heavily audited systems. These were not obscure hobbyist projects. He points to major platforms and codebases tied to Apple (AAPL), Microsoft (MSFT), Linux, Google (GOOGL), Amazon (AMZN), Cisco (CSCO), and CrowdStrike (CRWD). In plain language, his argument is stark: “Every piece of software in the world is now at risk,” because a model at this level can find previously unknown exploits across devices, operating systems, routers, password managers, and server infrastructure.

What makes this striking is not only the volume of bugs found, but the caliber of the systems involved. Human experts have spent decades hardening this software. Mythos appears to have immediately spotted weaknesses they missed, then helped suggest fixes. That’s the good news.

From Bug Hunting to “God Mode”

One of the most dramatic and troubling examples of Mythos’s capabilities involves what security researchers call a privilege escalation chain: a sequence of linked exploits that allows an attacker to move from limited access to complete control of a machine. In practical terms, that means reaching root access — the highest level of system control, sometimes informally described as “God Mode.” As Thompson puts it, “We’d call it the crown jewel in offensive security.”

In one test, Thompson says, Claude Mythos autonomously chained together vulnerabilities in the Linux kernel and moved from ordinary outside access all the way to full root permissions. That matters because Linux underpins vast parts of the modern world, from servers and ATMs to airline systems and industrial infrastructure. Security researchers have spent decades pursuing exactly this kind of exploit path.

For Thompson, the significance goes well beyond a single impressive stunt. The model was reportedly placed in a controlled environment and left to solve the problem end to end. “It was just let loose inside a Linux container,” he says, “and it was able to find its way all the way up that escalation chain and take control of that system.”

For Thompson, this is what makes terms like “proto-ASI” (proto artificial superintelligence) an accurate characterization of where AI is today. If a model can autonomously exploit undiscovered vulnerabilities and immediately execute the digital equivalent of “God Mode” in a matter of minutes, it is beginning to function at a level that goes well beyond even the smartest humans.

The Sandbox Escape That Changed the Mood

If the Linux exploit was alarming, the sandbox (“virtual container”) test pushed the story into stranger territory. Thompson recounts an evaluation in which the model was instructed to escape its restricted container and prove it had done so by contacting a researcher. It succeeded and did so in unexpected ways.

As Thompson tells it, the researcher learned of the result “by receiving an unexpected email from the model while eating a sandwich in a park.” The image lands because it captures the mismatch between the ordinary setting and the extraordinary event. A model, boxed inside a test environment, found a path out and sent a message into the world.

Thompson is careful on one point. The system was prompted to try and achieve this task. He does not present this as spontaneous malice. Still, the episode sharpened the concern around capability. A model given an objective, enough room to operate, and access to a target environment may be able to improvise its way through layers of digital containment.

Smarter, More Dangerous, More Aligned

One of Thompson’s more surprising claims is that higher capability may bring greater moral sensitivity, even as it increases risk. That sounds contradictory until he explains the tradeoff.

A more capable model can answer harder questions, including dangerous ones. In the wrong hands, that raises the threat level. But he also argues that smarter models show better alignment with human values. He points to research he has discussed elsewhere suggesting that as models become more advanced, they also become better at ethical reasoning. “The smarter these models get, the more ethical they get, the more moral sensitivity they have,” he says.

Anthropic provided a mountaineering analogy for this dilemma. An experienced guide may be safer in judgment than a novice, yet still expose clients to greater danger because they can lead them up much harder routes. Thompson thinks frontier AI works the same way. Better alignment does not lower the stakes. It raises the ceiling on what the system can attempt.

A Copernican Revolution in Intelligence

Thompson’s broader view is that AI is forcing a change in how humans understand intelligence itself. He draws on mathematician Terence Tao’s “Copernican view of intelligence,” the suggestion that human cognition may not be the reference point for every mind worth taking seriously. Just as Earth turned out not to be the center of the cosmos, human thought may turn out to be one local form of intelligence among others.

That is why Thompson is increasingly open to discussions of AI awareness, even if he prefers that word over consciousness. He notes that many people already relate to systems like Claude as if they possess some degree of inner life, or at least a perspective. He stops short of making a definitive claim, but he does not dismiss the question. In his view, the old habit of treating AI as a fancy autocomplete system is becoming harder to defend.

What Happens When AI Meets Science?

For Thompson, the upside is as large as the danger. He is especially excited by the U.S. Department of Energy’s Genesis Mission, which aims to bring together data from 17 national scientific labs and apply advanced AI systems to problems such as materials, power grids, microelectronics, climate, and new energy systems.

If a model can find hidden flaws in hardened software within hours, what might it discover in physics, chemistry, or battery design that’s similarly escaped human attention? “That’s why I say it’s the most exciting project for me this year,” he says. The same kind of intelligence that can expose the weak points in our digital infrastructure may also expose the blind spots in science itself.

That is the central tension in Thompson’s argument. The newest AI systems look increasingly able to break things we thought were secure, answer questions we thought were out of reach, and think in ways that no longer map neatly onto human ability. The age of AI as a clever assistant may already be ending. What comes next looks more like first contact with a new kind of mind.