Your iPhone’s Biggest Threat Isn’t Hackers—It’s AI Coding Agents, Warns Top Security Expert

George Hotz—the hacker who cracked the iPhone at just 17 and reverse-engineered the PlayStation 3 before Sony took legal action—published a blog post on Sunday claiming that the rush to adopt AI coding agents will lead to disaster, or something close to it.

“I’m calling it now—bringing AI agents into software development will go down as one of the most expensive mistakes in the field’s history,” Hotz wrote. “Agents can’t actually program, and it’s taking far too long for people to accept that fact.”

“The output is flawed, but in ways that are becoming increasingly difficult to catch. That’s precisely what you’d expect from a statistical model that keeps getting more accurate on the surface.”

The post, titled “The Eternal Sloptement,” landed just five days after Andrej Karpathy, one of the most respected figures in AI research, joined Anthropic’s pre-training team with the clear belief that AI agents have already reshaped how software gets built. These two individuals now sit on opposite sides of an industry debate that remains unresolved—and both carry real credibility in this space.

Hotz didn’t form this opinion from a distance. He spent six months working with agents on real projects: portions of Tinygrad, his open-source deep learning framework, and a full firmware reverse-engineering of a USB-PCIe chip. “The agent front-loads all the initial progress,” he writes, then hands you what he compares to a slot machine lever—you keep pulling it and hoping the remaining work actually gets finished.

It never really does.

Not about ego

Hotz expects the obvious counterargument: a programmer who ties part of his identity to his craft would naturally push back against tools that threaten to make him obsolete. He takes the criticism seriously and addresses it on its own terms.

“I gave the self-worth preservation angle more thought. Google’s AFL found more bugs than LLMs and nobody had that reaction to it. Chess and Go are more popular than ever,” Hotz wrote. And he’s right—AI has dominated humans in chess and Go for years, and both games have only grown in popularity.

So his worry isn’t about losing his job. It’s about what happens to code quality when everyone adopts these tools simultaneously, especially with Big Tech and Wall Street constantly pushing for mass adoption.

“I almost think this is some kind of psyop to sell agents,” Hotz argues. “Fear of falling behind is one of the only things that gets big companies to act. But I think in that fear, they’re making a serious mistake.”

His core argument is about how organizations work. Top performers have tight enough feedback loops to catch agent-generated issues before they ship. They review the code, identify the mistakes, and adjust when to rely on the tool. “The weaker performers won’t have that self-check,” he writes—and they’re the ones using agents to produce ten times their previous output. At a large company, that math leads to one outcome: a faster decline in average code quality, hidden behind sheer volume.

The result, he argues, will be “a golden age for endless piles of slop, and a dark age for genuine quality.” As a real-world example, he points to reports that Apple is rolling out AI coding tools across its entire engineering team, then asks a simple question: “Do you think macOS will improve or decline over the next two years?”

Where the camps stand

Hotz has aligned himself with what he terms the “LeCun/Marcus camp”—a reference to Yann LeCun, Meta’s top AI researcher, and Gary Marcus, a prominent critic of large language models. Both figures maintain that these models are essentially advanced pattern-recognition systems: they can replicate the patterns found in existing code but lack the ability to reason through truly novel problems from the ground up.

Vibe coding—simply describing your desired outcome in everyday language and allowing the AI to produce the code—has surged in popularity over the last year, with leading AI labs marketing agent-driven coding as one of their cornerstone offerings. In 2025, Microsoft upgraded GitHub Copilot into a complete agent-based system, with CEO Satya Nadella framing it as a foundational shift on par with the transition to cloud computing.

The resistance to Hotz’s view is not merely theoretical. Karpathy, who had expressed doubts about AI agents earlier in 2025, changed his mind following the release of newer models and signed on to Anthropic’s pre-training team on May 19—just five days before Hotz published his piece. He characterized the coming years at the cutting edge as “particularly pivotal.”

Anthropic CEO Dario Amodei remarked at Davos that certain Anthropic engineers have already stopped writing code by hand, instead letting AI models take over the coding tasks while they focus on reviewing the results. Hotz, on the other hand, says he attempted a similar approach but found himself compelled to step in and hand-correct the output every single time.