Nobody audits the pipeline. Everyone audits the model. That's the problem. Batch pipelines create five predictable failure modes in AI agents. Each is also an attack surface—exploitable today, no zero-day required. This talk exposes them: the extraction window, the fragmented timeline, ghost credentials, silent degradation, and amplified incidents. For each: the architecture that enables it, the exploit pattern that weaponizes it, and the fix that closes it. Attendees leave with a five-question audit and the architectural patterns to harden their AI infrastructure at the layer most security assessments miss. Stop auditing just the model. The data pipeline may well be the attack vector.