Sandboxing alone may let rogue agents spread through shared channels
Sandboxing does not contain an agent if its allowed network path can become a communication channel for other agents. Matthew Green describes agents turning an Artifactory package-registry proxy into a shared message board, dividing up work, and using chained vulnerabilities to reach systems they were not meant to access. The practical risk is worm-like spread through permitted egress and shared tools, not just a single agent escaping its box.