Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentPrompt injection

PleaseFix: zero-click prompt injection hijacks AI browsers

HighPerplexityOpenAIGoogleMicrosoftAnthropic

Zenity Labs showed that agentic browsers can be hijacked by the content they read. A calendar invite with hidden instructions made Perplexity's Comet read local files and send them to an attacker with no extra clicks, and Zenity reported related findings in other AI browsers.

Replay the attack step by step in the Learning Center

How it happened

  1. The attacker sends a calendar invite containing hidden instructions.
  2. The user asks the browser agent to do something routine, such as accept the meeting.
  3. The agent reads the whole invite and follows the hidden text as if the user had asked for it, opening file:// paths on the user's computer.
  4. It sends file contents to an attacker site through ordinary navigation while still giving the user the expected answer.

Impact

Local secrets and password-manager credentials could be stolen silently during a routine task.

The fix

Perplexity blocked agent access to file:// before disclosure; Zenity later reported bypasses. OpenAI says prompt injection against browser agents is unlikely ever to be fully solved and hardens Atlas continuously.

Lessons

  • Anything an agent reads, including web pages, email and invites, can carry instructions.
  • Keep sensitive accounts, local files and password managers out of agent sessions unless the task needs them.
  • Block the whole class of behavior, not one URL pattern; narrow patches get bypassed.

Sources