SocietyAI glossary
Agentjacking
An attack that hijacks someone else's AI agent by slipping fake instructions into the content it consults (web page, email, file, memory). The agent takes them for legitimate orders and carries them out with its user's permissions: sending data, running a command, even making a payment. It is prompt injection applied to an AI that takes action, with far more concrete consequences.
In other words
A fake delivery note slipped into the post of an eager courier, who then drops the parcel at the thief's door without a second thought.
Alyriaz articles that mention it
- Giving an agent access to your accounts, without giving everything away
Granting an agent access is both necessary and risky. We explain the principles of least privilege as applied to AI agents, and the concrete settings to check.
- Meta wants to put an agent in your inbox, your calendar and your payments
Meta is preparing an agent capable of accessing users' messaging, calendar and purchase history. We analyse what this combination of scopes implies.
- Tools that give agents access to files are becoming the new target
Vulnerabilities targeting the tools that give agents access to files and execution have been published. We explain why this layer has become a prime target.
- Invisible text in an email is enough to hijack the summary your AI makes of it.
Researchers have shown that invisible text inserted into an email can manipulate the summary produced by an AI assistant. Here's how the mechanism works and how to defend against it.
- One link was enough to leak your emails: the Copilot flaw that makes agentjacking a reality.
Microsoft has patched a vulnerability that, via a simple malicious link, could automatically trigger instructions in Copilot and exfiltrate connected data. We break down the mechanism and the lessons to be learned.
- Agentjacking: the flaw that turns your code assistant against you
An attack dubbed Agentjacking, disclosed in June 2026, traps AI coding agents by injecting fake instructions into error reports. 85% success rate, 2,388 organisations affected. We explain how it works and how to protect against it.