AI Agent Security (Course)
Agentjacking
Hijacking a coding agent by planting instructions in data it reads through a trusted integration, such as a public error-reporting key.
In more detail
The Cloud Security Alliance named the attack class after a technique that planted a fake diagnostic note in an error-tracking project using its public ingestion key. When a developer later asked an MCP-connected coding agent to check recent errors, the agent read the note as guidance and ran the command it suggested with the developer's own permissions. Every step used a legitimate action, so permission systems saw nothing wrong. The pattern applies to any service an agent trusts for output that strangers can write to.