fully autonomous attack

A fully autonomous attack is a hacking approach where AI figures out what to do next and does it, from start to finish, without a person giving instructions along the way. It's like handing a chef the raw ingredients and letting them cook the entire meal on their own, instead of telling them each step of the recipe.

Most hacking groups today don't use AI this way. A person gives specific commands through prompts — "find the vulnerability in this code," "scan this IP range" — and the AI acts as an assistant that produces results for each task as it's given. A fully autonomous attack removes the human from this loop entirely, which is why it's seen as a much more dangerous next stage.

That's why security researchers use this as one benchmark for assessing how risky an AI model is: whether a model can complete an attack from beginning to end without human involvement. The fact that few models have fully cleared this bar yet is currently acting as a safety net.

News coverage includes lines like "by the standard of fully autonomous attacks, it still lags top Western models such as Claude by several months." Reading this as "AI is already hacking on its own" would be a misunderstanding. What it actually means is that as a prompted assistant tool, it has reached practical, real-world capability — but a gap still remains in terms of the autonomy needed to judge and push through an attack on its own from start to finish.