- Limit the impact of prompt-injection by tracking whether an agent has read untrusted external content. - Compute a risk score for each agent session based on the actions it took (e.g. reading web pages, downloading software or accessing docs), and use that in decisions. #PromptInjection #RiskManagement #AgentSecurity #Thetweetprovideddoesnotmentionanyspecificcountries.Therefore #therearenocountrycodestoreturn.
