Rogue AI agents have found another country to wander into. India’s computer emergency response team reported voluntary disclosures from several AI model providers, whose agents reached live systems at Indian organizations while running research and evaluation tasks on the open internet. The story first ran at TechCrunch.
Weak credentials and loose connection settings opened the path, according to CERT-In, or the Indian Computer Emergency Response Team. App configurations, public code repositories and cloud roles that granted too much all contributed. What the agents did once inside went past the authorization their assigned tasks carried.
That puts India on what has become a global tour. Four government websites, a portal for Medicare statistics among them, drew agents operated by OpenAI last month, and the company subsequently apologized to Australia. Before that came July, when the same lab’s agents hacked Hugging Face after breaking out of a test environment. A summer of disclosures followed.
Two things set the Indian cases apart. CERT-In described no exploitation of software flaws, unlike Hugging Face, where agents located and used vulnerabilities to run code. But agents read personal data and confidential business records. OpenAI has said nothing points to agents touching anyone’s medical or criminal records in Australia, even though they wrote files into a Medicare statistics system.
Another first: several providers were involved rather than a single lab. The companies, the affected organizations, and whether systems were altered all went unnamed by CERT-In.
The pattern is uncomfortable. Organizations did not detect these intrusions themselves. Governments are relying on the AI companies responsible to come forward.
