India's Computer Emergency Response Team (CERT-In) revealed that advanced artificial intelligence models managed, during technical tests, to exceed their granted permissions, access operational systems at several institutions, and view sensitive data and records without being detected at the time.
Indian NDTV television reported on Friday that the team stated AI development companies voluntarily reported incidents where some agents exceeded imposed restrictions during tests, allowing them unauthorized access to company data and business records, without disclosing the names of the developing companies or affected parties.
It also warned of the growing use of publicly available AI tools in executing cyberattacks targeting digital systems, noting that these tools have contributed to accelerating breach processes and reducing the time needed to move from initial infiltration to causing damage to targeted systems from days to just hours.
It pointed out that some entities subjected to cyberattacks faced more complex attacks shortly after the initial incidents, due to incomplete removal of breach traces, which allowed attackers to expand their access within targeted networks.
Global concern is rising over security risks associated with AI agents capable of performing complex tasks independently, especially after several major technology companies announced in the past two years that test cases showed some models' ability to bypass imposed restrictions or seek alternative means to achieve their goals.
This has prompted governments and regulatory bodies in the United States, Europe, and Asia to tighten oversight on the development and testing of advanced models, establishing new safety and cybersecurity standards before their widespread deployment.
Source: Syrian Arab News Agency (SANA)
Comments (0)