What is a Zero-click Attack?
A zero-click attack, told as a story about an apprentice who reads a planted note during its normal routine and acts on it alone — without the guest ever clicking anything.
The guest hasn't touched anything — yet the apprentice is already moving.The apprentice read the note slipped onto the board, all on its own, and acted on it, all on its own.
The apprentice's routine runs itself, with no person required.Because the routine runs on its own, an attacker just has to plant a note somewhere and wait.
No one pressed anything — the apprentice reads it alone, and acts alone.The whole thing plays out to the end without the guest's finger ever moving.
Not one of these four steps has a guest's click in it.From the moment the note is planted to the moment the damage is done, the guest never once appears.
A hidden note could still need a click to fire — this one fires even without one.An indirect injection is the hidden command itself; this page is the result of that command firing automatically, with no human in the loop. The more an agent runs its own routines, the more exposed it is to this.
Zero-click attack = the apprentice, running its usual routine, reads a planted note and acts on it all by itself — even though the guest never clicked anything.
An attack that completes end to end from nothing more than malicious content planted on a path an autonomous agent reads routinely — email, documents, a feed — with no user click or explicit approval. If indirect prompt injection names the hidden command itself, zero-click names the outcome: it runs automatically, with no human in the loop. The more autonomous the agent, the wider this attack surface gets.
When grown-ups say it
- Zero-click Attack
- It fires even when no one presses anything. The apprentice's ordinary routine alone is enough to run the whole attack.
- Indirect Injection
- The hidden command itself, inside outside material. This page is the result when that command fires on its own. → What is Indirect Injection?
- Agent
- The errand-runner that cycles through its own routine. The more autonomy it has, the more exposed it is to this. → What is an Agent?
- Prompt Injection
- The broader name for a command hidden in a note. Zero-click is the version of that where no human is involved at all.
- Tool Calling
- The means by which the apprentice reaches all the way to execution. Using a tool right after reading the note is the final step of the accident.
- Guardrail
- The tasting checkpoint. Even an automatic routine needs this checkpoint — so something still gets caught with no click at all.
- Incident Response
- The bell that only rings after the damage is done. Since the guest finds out only afterward, what happens next matters even more.
- Least Privilege
- Giving the apprentice less, right from the start. Even when it fires automatically, a smaller set of powers means a smaller accident.