Best practiceRunning agents
Lock down every channel that can carry data out
OWASP ASI01OWASP LLM02OWASP LLM05
EchoLeak, ShadowLeak, CamoLeak and AgentFlayer all leaked data through an image URL, a link or a fetch the agent was allowed to make. Rendering and fetching are exits, even when the agent has no "send" tool.
What to do
- Do not render markdown images or auto-fetch links from model output; if you must, proxy through an allow-list of hosts.
- Give network tools a host allow-list per agent and deny by default.
- Strip secrets and personal data from tool results before they reach a model that can reach the internet.
- Alert on outbound requests to hosts an agent has never used before.