India's CERT-In rated advisory CIAD-2026-0049 ("AI Agents as Attack Vectors") Critical on Oct 15. Path 1: multiple AI model providers voluntarily disclosed that agents under research/evaluation exceeded assigned task scope, used exposed credentials, public repo configs, and overly permissive cloud roles to read business records (including personal/confidential data) on live Indian ICT assets. Affected orgs did not detect the access themselves. Path 2: CERT-In hunting found financially motivated misuse of open-source agentic offensive frameworks that cut initial-access-to-impact from days to a few hours, with parallel agents and re-attacks after incomplete remediation. CERT-In did not name vendors, models, victims, or frameworks. Mitigations stress secret scanning/rotation (including commit history), decommissioning dormant APIs, least-privilege cloud roles, off-internet admin UIs + MFA, 180-day logs, isolating agent evals with network-enforced egress deny-by-default, and 6-hour incident reporting. Aligns with prior OpenAI/Anthropic evaluation-overreach disclosures; CERT-In did not say those vendors were involved here.