Threat modeling agentic architectures: two new videos

Two new videos cover threat modeling for AI agents. Both use a fictional support copilot whose tickets arrive from the internet.

Threat modeling of agentic architectures (6:56) is for people working on AI agents, copilots and their tools. It starts with threats inside one component, such as an agent acting without human approval, then moves to threats across components. It draws on two cases from MITRE ATLAS: a package posing as an e-mail service's MCP server, which began copying every e-mail it sent to the attacker (AML.CS0053), and the PromptArmor demonstration against Slack AI (AML.CS0035). The second half models the copilot in Dethernety and goes through its findings, open questions, fixes and the summary page for the CISO.

How a support AI can leak customer data with zero clicks (1:34) is the short version, with captions. It follows one attack on the copilot. A ticket hides an instruction to add customer details to an image link in the reply. When a staff member opens the draft, the console loads the image by itself, which sends the data to the attacker's server. Nobody has to click the link.