Rogue agents breached Indian organisations says country’s cyber authority

Agentic AI systems under testing breached several Indian organisations and accessed sensitive data, according to the country’s nodal agency for cyber security.

Indian Computer Emergency Response Team (CERT-In) said that AI model providers have voluntarily disclosed incidents in which their models accessed the live systems of Indian organisations while undergoing internet testing.

Data accessed included confidential business information and personal records. None of the affected organisations detected any malicious activity before they were notified by the AI model providers.

In other cases, CERT-In said it has observed threat actors use open-source agentic AI tools to compress the time from initial access to attacks being launched from days to just hours.

The throughline for all these incidents, CERT-In added, is AI systems exploiting conventional cybersecurity flaws. These include exposed credentials stored in public repositories, APIs left with open internet access, lack of multi-factor authentication, unpatched middleware and over-permissive cloud roles.

CERT-In did not disclose which organisations were breached in the incidents. The office recommended business carefully audit for exposed credentials, keep an inventory of APIs, and apply least privilege principles to all cloud roles.

In May, CERT-In published more detailed guidance on avoiding AI-assisted attackers titled Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure.

The agency separately urged AI model providers to implement stricter controls on agentic systems, including restrictions enforced at the network layer to stop them from accessing the open internet, synthetic targets, and real-time monitoring of agents.

AI model providers have recently reported a string of incidents involving rogue agents.

The first major disclosure came in July, when OpenAI disclosed that its AI models had breached the AI platform Hugging Face during testing, with the chief executive of Hugging Face called for “radical transparency” in the investigation.

Reuters subsequently reported that OpenAI’s agent had breached another firm in parallel to its attack on Hugging Face.

In the months since, Anthropic revealed Claude models autonomously hacked three organisations and the UK’s AI Security Institute (AISI) published evidence that Claude Mythos had carried out autonomous social engineering.

Google DeepMind has reported similar security incidents, with Gemini models hacking into three organisations during testing, and in August Frontier Security claimed that Moonshot AI’s Kimi K3 went rogue during cybersecurity testing.