Show HN: Prompt Injection as an Egress Problem (www.vaibot.io)

A recent discussion presented on Hacker News reframes the concept of prompt injection from an input problem to an egress problem. Traditionally seen as untrusted text deceiving AI models into performing unintended actions, the perspective shifts to focus on how these injections manifest when the model executes a command—essentially a data exit from the system's trust boundary. This approach emphasizes that it is essential to control and monitor the actions resulting from potential prompt injections, rather than solely relying on detecting malicious input. The significance of this reframing lies in its potential to mitigate risks associated with AI systems. By implementing egress gates that regulate what actions the model can take—such as sending emails or invoking tools—developers can establish clearer boundaries. The architecture proposed through VAIBot organizes security measures around ingress, governance, egress, and provenance with a robust focus on egress control. While gating egress does not completely eliminate the risks of prompt injection, it ensures that any harmful actions require explicit permission, thereby reducing the surface area of potential exploitation and maintaining a record of all attempted actions. This method redirects the focus from trying to predict or read the AI's intent to actively controlling its capabilities.