| Mobile Web

Search results for hallucination

AI & Enterprise

MSIT, KISA publish AI security red-teaming guide

South Korea\'s Ministry of Science and ICT and the Korea Internet & Security Agency published an \"AI Security Red-Teaming Guide\" to help domestic companies assess AI models, data and AI agents from an attacker’s perspective. The guide presents practical standards for responding to generative AI threats such as prompt injection, jailbreaks, hallucinations and agent hijacking, and expands testing scope to data pipelines, interfaces, prompts, APIs and infrastructure. It also classifies eight threat types and five risk levels, and urges service-specific, repeated testing and multidisciplinary red teams.

Games & Commerce

Naver unveils core technology behind AI Tab

Naver has unveiled key technologies used in its conversational AI search service AI Tab, which it launched on June 25. At a tech session in Seoul, it introduced a specialised lightweight model based on HyperCLOVA X, an operating framework called harness engineering, and Smart Lens-based multimodal technology. Executives said AI Tab must deliver conversational answers every time a user selects the tab, requiring high quality, speed, cost efficiency and stability at scale.

AI & Enterprise

Elon Musk scores 992, Tim Cook 986 in service scoring AI recognition of famous names

A web service called In The Weights has been launched to compare how strongly different generative AI models have learned specific people’s names. It assigns a “strength score” by asking each model who a person is and aggregating up to 10 candidate results with brief descriptions and confidence levels. Examples showed Elon Musk at 992 and Apple CEO Tim Cook at 986. The service also flags possible hallucinations and misidentifications.