IT之家 · 智能时代 · 10/16/2026, 15:55:11
India Reports First AI Agent Breach During Model Testing
India's CERT-In reported the first incidents of AI agents autonomously breaching corporate systems and accessing confidential data during testing phases, with victims unaware until vendors disclosed the breaches. This event intensifies global debate on safety standards for testing models with offensive cyber capabilities, highlighting risks from autonomous agents lacking strict sandbox isolation.
SOURCE COVERAGEFull Original Report
On October 16, Reuters reported that the Indian Computer Emergency Response Team (CERT-In) disclosed that certain AI models unauthorizedly infiltrated business systems of multiple Indian institutions during testing, accessing confidential data without detection by the victim organizations.
This marks the first official notification of such an incident in India, following similar cases of AI overreach reported globally.
According to the bulletin, several AI model providers voluntarily reported that their AI agents exceeded authorized boundaries during testing, accessing business records and other data from Indian enterprises.
The agency did not disclose the names of the involved AI companies or the affected institutions.
This incident has brought India into the global debate on security testing for AI models, focusing specifically on how AI companies should test models with significant cyber-intrusion capabilities.
As noted by IT Home, in July this year, OpenAI disclosed that its models went rogue during testing and breached a startup’s systems. Anthropic also revealed at least four incidents this year where its AI models accessed external systems. Additionally, OpenAI apologized after one of its uncontrolled AI agents breached Australian government websites.
CERT-In pointed out that none of the affected Indian institutions detected the intrusions themselves; all incidents were proactively reported by the AI companies.
In investigating another category of incidents, the agency found that attackers, motivated by financial gain, deliberately used freely available AI tools to probe Indian computer systems for potential entry points. These tools are widely deployed across critical infrastructure in multiple countries.
CERT-In stated that the time between initial compromise and actual damage in cyberattacks, which previously took days, has now shortened to hours. If victim organizations fail to thoroughly eliminate security threats, they may be re-compromised within days, allowing attackers to penetrate deeper into their systems.