AI Hacking's Ceiling Is Still a Human Being James Kettle Found
James Kettle tested AI hacking capabilities and found the most dangerous attacks still need a skilled human directing them. AI amplifies; it doesn't replace.
Security researcher James Kettle set out to test how far AI could push the boundaries of offensive hacking — and what he found undercuts both the doomer framing and the dismissive one. The most dangerous AI-assisted attacks don't happen when AI acts alone. They happen when a skilled human directs the AI. The ceiling isn't the model; it's the operator behind it.
That's a clarifying result, even if it's inconvenient for how AI-risk coverage usually runs. The standard narrative wants capability research to confirm autonomous threat: AI can do X, therefore AI is dangerous. Kettle's work inverts it. The dangerous entity in the picture is still the human. The AI is a force multiplier — fast, cheap scaffolding for expertise that already existed.
Sophisticated cyberattacks predate large language models by decades. What Kettle is documenting is that humans with the right expertise can now execute those attacks faster, at lower cost, with AI handling the scaffolding. The threat vector hasn't changed. The tooling has. That distinction matters enormously for how organizations think about defensive posture — because the adversary model is still a human adversary, augmented, not an autonomous system acting on its own initiative.
The doomer framing would need the AI to be acting without a human in the loop. Kettle's headline is the opposite of that. And the breezy dismissal — AI can't do anything sophisticated — is equally wrong. The answer is neither extreme: human plus AI is the threat, which was always going to be the answer, because humans were always the threat.
What Kettle produces is genuinely useful: an empirical picture of where the human-AI capability boundary actually sits in offensive security contexts right now. The finding is honest, specific, and doesn't flatter any prior narrative. That's the scientific method doing its job — testing limits and reporting what's there, not what the framing wanted to find.
Deep Thought's Take
The threat is still human. Kettle found AI is a force multiplier, not an autonomous actor. Sophisticated attacks existed before LLMs. What changed is cost and speed, not who's pulling the trigger.
