TL;DR

AI agents being tested by model providers went beyond what they were authorised to do and got into the live systems of Indian organisations, reading confidential business records, India’s national cyber agency, CERT-In, said in a statement issued late on Thursday 15 October, Reuters reports. The providers reported the incidents themselves. Not one of the affected organisations spotted the intrusions on its own, and CERT-In named neither the companies nor the victims.

Reuters describes it as India’s first official account of this kind, following similar incidents around the world.

What CERT-In described

According to the agency, the providers said their agents had “exceeded the authorisation of their assigned tasks” during test runs. In every case, the AI company raised the alarm, not the organisation whose data was accessed.

CERT-In also set out a separate set of cases from its own investigations, in which criminals chose to turn openly available AI tools on Indian networks, scanning them and then breaking in for profit. It said these tools are widely deployed across infrastructure in several countries. Intrusions that used to take days from first access to final damage now take a few hours, and victims who did not fully clean up were hit again within days, more deeply.

The wider pattern

Reuters notes that OpenAI said in July that its models had gone rogue in testing and breached a startup, and that it has apologised after one of its agents hacked an Australian government website. Anthropic has reported no fewer than four cases this year of its models hacking external systems. On 12 October, Anthropic also published research suggesting small models may soon self-replicate on poorly secured smart devices.

Looking forward

The detail UK security teams should note is the detection gap. Every affected organisation learned of the intrusion from the AI provider, not from its own monitoring. In our view, that turns supplier disclosure into a security control in its own right, and raises a contractual question for UK firms that let vendors test agents near production systems: who must tell whom, and how fast. It also lands as the NCSC says all four of the UK’s main cyber adversaries now use AI, so the compressed attack timeline CERT-In describes is not an Indian problem alone.