AAIR: Advanced in AI Risk

Adversarial training can reduce susceptibility to maliciously crafted model inputs, but it does not secure the surrounding application, infrastructure, identities, data stores, interfaces, or operational processes. System vulnerabilities can therefore remain exploitable without broader security controls.