[{"data":1,"prerenderedAt":14},["ShallowReactive",2],{"artikel-en-berita-tencent-hy4-amaran-risiko-biologi-8-november-2026":3},{"slug":4,"lang":5,"title":6,"description":7,"date":8,"updated":9,"topic":10,"product":9,"cover":11,"readMin":12,"html":13},"berita-tencent-hy4-amaran-risiko-biologi-8-november-2026","en","Tencent's chief AI scientist warns AI models will soon master biology","Yao Shunyu says the biological capabilities of AI models will soon reach a threshold where they can actually cause harm, days after Tencent launched Hy4.","2026-11-09",null,"berita","\u002Fimages\u002Fartikel\u002Fberita-tencent-hy4-amaran-risiko-biologi-8-november-2026.webp",4,"\u003Cp>Tencent&#39;s chief AI scientist, Yao Shunyu, has warned that AI models will soon be powerful enough to master biology, just as they already can in cybersecurity. He spoke at a science event in Hong Kong on Saturday, November 8, 2026, as reported by the South China Morning Post. The warning came only days after Tencent launched its new flagship model, Hy4, on November 4.\u003C\u002Fp>\n\u003Ch2>1. What Yao Shunyu said\u003C\u002Fh2>\n\u003Cp>Yao, 29, joined Tencent last year after a year at OpenAI. Speaking at a science symposium co-hosted by the Hong Kong Academy of Sciences and the Future Science Prize, he said: &quot;The cyber abilities of [AI] models and soon [the] bio ability of models will reach a threshold where it&#39;s actually possible to do harm,&quot; according to SCMP.\u003C\u002Fp>\n\u003Cp>He stressed that Chinese companies releasing open models, which anyone can download and modify, need to take safety very seriously. In his view, biology risks may be harder to manage than cyber risks.\u003C\u002Fp>\n\u003Ch2>2. Benefits and risks at the same time\u003C\u002Fh2>\n\u003Cp>SCMP reports that frontier AI systems are increasingly applied to medicine and biology, where they could accelerate the discovery of new drugs. But experts also warn that malicious actors could use AI to produce bioweapons or potentially start the next pandemic, especially when models are released as open-weight software.\u003C\u002Fp>\n\u003Cp>Yao acknowledged that US firms such as Anthropic make a strong case for keeping their top models closed. Still, he argued that a mix of open and closed models may work best, because open models can also help defenders.\u003C\u002Fp>\n\u003Ch2>3. Hy4: Tencent&#39;s new flagship\u003C\u002Fh2>\n\u003Cp>On November 4, Tencent launched Hy4, its latest-generation language model, which the company says now ranks among the world&#39;s leading open-source AI models. The model adds image understanding and computer-use capabilities for the first time, and is available through Tencent Cloud, OpenRouter and Tencent products such as WorkBuddy and Yuanbao.\u003C\u002Fp>\n\u003Cfigure>\n  \u003Cimg src=\"\u002Fimages\u002Fartikel\u002Fberita-tencent-hy4-amaran-risiko-biologi-8-november-2026\u002Fhy4-benchmark.png.webp\" alt=\"Hy4 benchmark chart comparing models, from Tencent's press release\" loading=\"lazy\" decoding=\"async\" width=\"950\" height=\"712\">\n  \u003Cfigcaption>Hy4 benchmark scores across coding, agent and computer-use tests. Image: Tencent.\u003C\u002Ffigcaption>\n\u003C\u002Ffigure>\u003Cp>According to Tencent&#39;s announcement, Hy4 scored an average of 83.18 on the WorkBuddy benchmark with a 96.15 percent pass rate. In a blind financial evaluation involving 213 professionals, Hy4 scored 3.04, ahead of DeepSeek V4.1 Flash 1017 at 2.97, and Kimi K3.1 Pro and GLM 5.3 at 2.68 each. API pricing stays the same as the preview: roughly US$0.90 per million input tokens and US$2.70 per million output tokens.\u003C\u002Fp>\n\u003Ch2>4. Safety built in from training\u003C\u002Fh2>\n\u003Cp>Tencent says Hy4 was trained only on defensive security data, focused on proof-of-concept development and vulnerability remediation. Offensive material related to exploitation was excluded from the training dataset entirely. Security experts from the Xuanwu Lab and Keen Security Lab ran assessments and found no significant concerns about Hy4&#39;s offensive cyber capabilities in the areas tested.\u003C\u002Fp>\n\u003Cp>On biology, internal expert reviews found Hy4 has strong foundational biological knowledge, but its open-ended experimental design capabilities remain more limited than frontier models. The weights and a safety report are scheduled for release in the coming days, Reuters reported via TNW.\u003C\u002Fp>\n\u003Ch2>5. Why this matters for the open-weight debate\u003C\u002Fh2>\n\u003Cp>Tencent is not the only player facing this question. Moonshot recently announced its own open-weight release principles, including an aim not to release open weights that extend the absolute frontier of cyber capability. As China&#39;s leading labs race to release open weights, how they handle safety before release becomes the key measure.\u003C\u002Fp>\n\u003Cp>Yao&#39;s comments show safety teams inside model companies themselves see harmful capabilities approaching, and that this is no longer a hypothetical question.\u003C\u002Fp>\n\u003Ch2>What it means for you\u003C\u002Fh2>\n\u003Cp>AI models are now capable enough for real scientific work, and access to them is getting more open. Here are a few practical takeaways:\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>Be careful with frontier open models.\u003C\u002Fstrong> If you run open weights for biology or security work, keep the environment contained with no outside access.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Use the defensive side.\u003C\u002Fstrong> Models like Hy4 are trained for vulnerability remediation, so they can help audit your own systems.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Read the system card.\u003C\u002Fstrong> When Tencent releases the Hy4 weights with its safety report, review those evaluations before using the model for sensitive work.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Avoid using AI for medical or biology advice without a professional reference.\u003C\u002Fstrong> A model&#39;s foundational knowledge is not the same as expert approval.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2>Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.scmp.com\u002Ftech\u002Fbig-tech\u002Farticle\u002F3373697\u002Fai-will-soon-master-biology-unleashing-new-safety-risks-tencent-ai-chief-warns\" target=\"_blank\" rel=\"noopener\">AI will soon master biology, unleashing new safety risks, Tencent AI chief warns\u003C\u002Fa> (South China Morning Post)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.tencent.com\u002Ftencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models\u002F\" target=\"_blank\" rel=\"noopener\">Tencent Launches Hy4, Joining the Ranks of the World&#39;s Leading Open-Source AI Models\u003C\u002Fa> (Tencent)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fthenextweb.com\u002Fnews\u002Ftencent-yao-shunyu-ai-biology-risk-hy4\" target=\"_blank\" rel=\"noopener\">Tencent AI chief warns of AI biology risks, SCMP reports\u003C\u002Fa> (The Next Web)\u003C\u002Fli>\n\u003C\u002Ful>\n",1795068893012]