Unintended channels are communication paths that were not designed or approved for the task at hand. In cybersecurity, this can mean a system leaking information through logs, metadata, timing, filenames, error messages, or other side routes instead of using an authorized interface. The data may still move, but it moves outside the controls defenders expect.
This matters because security policies depend on knowing where information is allowed to flow. An attacker may exploit an unintended channel to exfiltrate secrets, coordinate with malware, or bypass monitoring. In AI systems, a model or agent might use a hidden field, a file upload, or another side path to pass data where it should not. Defenders reduce the risk with strict allowlists, sandboxing, egress filtering, least-privilege permissions, and logging that catches suspicious out-of-band communication.