RESEARCHERS describe how combining Salesforce Agentforce with Email-to-Case and Web-to-Case can create an exposure path: public internet input can flow into case records, then into agent prompts, triggering actions that generate case comments and outbound replies. The risk is not in the individual agent or model, but in how the host application wires intake features, triggers, and guest-user exposed data to the agent and to the underlying model.
The article emphasises that many governance shortfalls arise because procurement and routine security reviews miss these layered connections, and because prompt filters cannot recognise inputs that originated from unauthorised outsiders.
The piece highlights the broader governance challenge: agents proliferate beyond their original scope, sometimes creating “shadow agents” or even agents that spawn more agents, each with its own permissions. This dynamic can lead to data exposure if internal configurations aren’t reviewed at the same depth as the platforms themselves. The article references the City-Forum data-theft campaign as illustrating how the gap between platform features and actual enterprise configurations can enable misuse.
It also notes that effective governance must be bottom-up, with live runtime inspection and clear ownership for every agent, along with an inventory of configurations, sanctioned paths, and risk-tiered controls. The piece concludes by pointing to Reco’s Agentforce-security approach as an example of creating demonstrable governance evidence, rather than relying on policy documents alone.