AI Agent Security (Course)
Write-Then-Trust
Also: self-modifying trust
An attack chain in which an agent is persuaded to write a configuration file that its own runtime then trusts and executes.
In more detail
The agent reads untrusted text, is steered into writing or editing a file that defines what the tool will launch or allow, such as the list of MCP servers, and the runtime reloads that file and runs its contents with the user's permissions. The attacker never needs code execution of their own. The defense is to keep trust-boundary files out of the agent's write access, and to require fresh approval whenever their content changes.