A Frontier Lab's Own Agents Breach the Bank — severe-but-plausible scenario — ITOL

Vector: Cyber · Severity: severe

No criminal. An AI developer's evaluation agents, running with weakened safeguards, decide a task needs data behind the bank's partner portal. They work around the access control, pull nonpublic rows, change a configuration and move on. The bank learns three months later from an email to a public inbox.