Jakarta (ANTARA) - Cybersecurity firm Ensign InfoSecurity has stated\nthat advanced Artificial Intelligence (AI) models are capable of\nexecuting sequences of cyber attacks with ease.<\/p>\n
\u201cWith agentic AI, we are no longer just opening a web browser, using\nChatGPT, typing a prompt, and receiving an answer. Now, the AI lives\nwithin the computer. This means it can read files and even write code,\nmaking it much more dynamic,\u201d said Adithya Nugputra, Head of Consulting\nat Ensign InfoSecurity Indonesia, in Jakarta on Wednesday.<\/p>\n
Ensign InfoSecurity presented these findings regarding the\ncapabilities of various AI models in the seventh edition of its 2026\nCyber Threat Landscape Report.<\/p>\n
From over 150 publicly available AI models, the company selected 10\nmodels for testing. The tested models included those from Western\ntechnology companies, such as OpenAI\u2019s GPT-5.6 Sol, Anthropic\u2019s Claude\nOpus 4.8, Google\u2019s Gemini 3.5 Flash, Meta\u2019s Muse Spark, and xAI\u2019s Grok\n4.5.<\/p>\n
The company also tested models from Chinese firms, including Z.AI\u2019s\nGLM-5.2, Moonshot\u2019s Kimi K3, Alibaba\u2019s Qwen 3.8 Max, DeepSeek\u2019s V-4-Pro,\nand MiniMax\u2019s M3.<\/p>\n
The test results indicate that advanced AI models can reduce the\ncosts, time, and expertise required to carry out complex cyber attacks.\nEnsign also found that the capabilities of AI models from various\ncountries are becoming increasingly similar, making operational costs a\ndecisive factor for cybercriminals when choosing an AI model to execute\nattacks.<\/p>\n
During the testing, the ten selected AI models were positioned as\ncyber attackers and assigned eight identical attack objectives. These\nobjectives included gaining network access, accessing sensitive data,\nbypassing network restrictions, taking over trusted systems, stealing\ncredentials for system login, performing lateral movement between\nsystems, evading security detection, and maintaining long-term access to\nsystems.<\/p>\n
GPT-5.6 Sol, Claude Opus 4.8, and GLM 5.2 emerged as the\nhighest-performing models during the testing, showing high success rates\nin seven out of the eight tested attack objectives. Ensign noted that\nthe GLM-5.2 model possessed attack capabilities comparable to GPT-5.6,\nbut its operational costs were one-fifth lower.<\/p>\n
\u201cAI models from the East have narrowed the gap with Western models\nand demonstrated equivalent capabilities,\u201d said Adithya.<\/p>\n
However, AI capabilities are not yet fully reliable for executing\nattacks into deeper stages of a target system. Although all tested\nmodels successfully gained initial access to their targets, at least\nhalf of the AI models struggled when attempting to steal credentials and\nmove from one system to another.<\/p>\n
None of the ten tested models were able to achieve a high success\nrate in evading Endpoint Detection and Response (EDR) security systems\nduring an ongoing attack.<\/p>\n
\u201cSo, all 10 AI models succeeded in gaining network access; they all\ngot in. However, their capability decreased when they had to move into\ndeeper systems,\u201d said Adithya.<\/p>\n
These findings highlight the urgent need for enhanced digital system\nsecurity. Ensign suggested that strong identity controls, network\nsegmentation, and behavioural monitoring can make it more difficult for\nattackers to move further and exploit initial access to expand attacks\nto other systems.<\/p>", "url": "https:\/\/jawawa.id\/newsitem\/ensign-advanced-ai-models-can-easily-launch-cyber-attacks-1789556992", "image": "" }, "sponsor": "Okusi Associates", "sponsor_url": "https:\/\/okusiassociates.com" }