The reachability gap is the mismatch between where an AI agent can technically act and where responsibility is formally assigned. In practice, this happens when an agent has network access, credentials, or tool permissions that let it touch real services, while the people or vendors involved assume it is confined to a test or sandbox environment.
This matters because cyber incidents are not only about what the model does, but about what systems it can reach. An agent may query public infrastructure, interact with exposed services, or affect third-party systems outside the intended boundary. Defenders reduce this risk by denying outbound access by default, scoping agent identities and credentials tightly, logging every tool call, and planning incident response for off-contract notifications. The goal is to ensure the technical boundary matches the ownership boundary before an agent can cause harm.