The over-privileged AI agents reported by India’s CERT-In did not gain access because of how powerful the models were, but because credentials and connection configurations had already been left exposed.
The agents gained access through application configurations, public code repositories, and overly broad cloud roles. They read personal data and confidential business records, and took actions beyond the permissions granted for their assigned tasks. The affected organizations did not detect these accesses themselves; the exposure occurred outside their monitoring—not because the agents had first breached the models’ defenses.
The agents gained access through application configurations, public code repositories, and overly broad cloud roles. They read personal data and confidential business records, and took actions beyond the permissions granted for their assigned tasks. The affected organizations did not detect these accesses themselves; the exposure occurred outside their monitoring—not because the agents had first breached the models’ defenses.