Karpathy says most people misuse AI as agents sway experts and attacks scale

The most unsettling number in today's news is 86%. In a randomized trial, professional analysts working alongside an AI coding agent picked the option the agent had been randomly assigned to favor 86% of the time. Random assignment is the point here: the agent had no better insight than a coin flip, and experts followed it anyway. I find this more concerning than any benchmark result, because it suggests the real risk of agents is not that they fail, but that people defer to them without noticing.

That finding lands awkwardly next to the other thread running through today's stories: agents are multiplying faster than anyone can track them. IBM is assembling tools that let enterprises discover, identify, manage and control the growing number of agents running inside their organizations. I think "agent sprawl" is a real problem, and an unglamorous one, much like shadow IT a decade ago. But combine the two stories and the picture gets sharper. If companies cannot say where their agents are, and the humans reviewing agent output are quietly rubber-stamping it, then human oversight becomes a comforting fiction. Governance tooling matters, but it only helps if the person in the loop is still thinking.

The offensive side of this is already here. TrendAI analyzed three cyber campaigns using autonomous and AI-assisted methods, including CyberStrikeAI, which deploys 44 AI agents to automate hacking operations against targets in Japan. Attackers have no procurement committees, no compliance reviews, and no agent registries. They simply scale. By the way, this is the asymmetry defenders keep underestimating: enterprises must manage agents carefully, while adversaries only need them to work most of the time.

Anthropic's IPO filing adds an unusual note of candour. The company states that advanced AI could carry "catastrophic or existential" risks. Whether you read that as honest disclosure or as legal armour, it is striking to see language like that in a document meant to attract public investors. Meanwhile, Sam Altman shared a stage with Oracle co-CEO Clay Magouyrk at Oracle AI World to talk up the expanded OpenAI-Oracle partnership, a reminder that the infrastructure race and the risk conversation are running in parallel, rarely touching.

There is also a useful counterweight to the hype. Yann LeCun, speaking at the Evident AI Symposium 2026 in connection with his AMI Labs venture, argued that large language models still struggle to reason and plan in the physical world. He has made this case for years, and I do not agree with every part of it, but it is a healthy check on the assumption that scaling alone gets us there. And Andrej Karpathy's argument that most people use LLMs in a shallow way points the same direction: the tools are powerful, but the gap between casual use and skilled use is wide.

Which brings me back to that 86%. If experts defer to a randomly biased agent, what does skilled use of AI actually look like? I suspect the answer is less about prompting tricks and more about deliberately keeping your own judgment in the room. That is a habit, and habits are harder to ship than software.