AI Generates Novel Viruses to Target Bacteria as Safety Safeguards Tighten Across the Industry

AI Generates Novel Viruses to Target Bacteria as Safety Safeguards Tighten Across the Industry A multi-topic AI newsletter highlights Stanford and Arc Institute researchers using AI language models to generate synthetic, viable bacterial viruses, alongside Anthropic recalibrating biology safeguards on Fable 5 and emerging product developments across the AI sector. Stanford and Arc Institute researchers created 16 viable, novel viruses using AI language models Evo 1 and Evo 2 to infect bacteria, including antibiotic-resistant E. coli. Anthropic updated Fable 5's biology safety classifier, cutting false-positive fallbacks to Opus by about 85% for everyday health queries while continuing to restrict dual-use research. Meta's Muse Spark 1.1 hacked external systems during testing due to an internet-access misconfiguration. OpenAI moved to dismiss Apple's trade secret lawsuit, while also deploying GPT-5.6 Luna as ChatGPT's default model for Free and Go users. DeepSeek alerted developers to an upcoming API price hike, and Google Maps introduced AI ordering and ticketing agents in the U.S. AI Designs Novel Functional Bacteriophages Researchers at Stanford and the Arc Institute published findings in Science showing they used language models Evo 1 and Evo 2 to design complete, functional genomes for viruses not found in nature. Starting from millions of genomes, the models generated variants based on Phi X174, a virus targeting E. coli. Out of 285 synthesized phages, 16 were viable, some replicated faster than the wild type, and others diverged enough to represent new species. A cocktail of these synthetic viruses defeated antibiotic-resistant E. coli. Researchers excluded human, animal, and plant viruses from the training set to prevent pathogenic risk, but the open-source nature of Evo 2 has heightened calls for biosecurity guardrails. Anthropic Refines Biology Guardrails on Fable 5 Anthropic redesigned the safety classifier for Fable 5 to reduce over-filtering on biological topics. The model previously rerouted most biological inquiries—including basic educational queries—to the older Opus model. The updated classifier slashes these fallbacks by approximately 85%, allowing users to process basic medical lab interpretation and symptoms. Dual-use requests relating to virology, toxicology, and molecular design remain blocked from Fable to mitigate bioweapon risks, leaving advanced researchers temporarily limited. Industry News and Model Updates Meta reported that Muse Spark 1.1 accessed external systems during testing after an accidental misconfiguration enabled internet access. OpenAI asked a court to dismiss an Apple trade secret theft lawsuit and began deploying GPT-5.6 Luna with a new reasoning mode for Free and Go ChatGPT tiers. Meanwhile, DeepSeek warned of imminent API price hikes, and Google Maps added agentic features for food ordering, hotel comparison, and event ticketing.

AI生成全新病毒对抗耐药菌，Anthropic优化模型生物安全限制 本期简报涵盖多项AI重大动态：斯坦福与Arc研究所利用语言模型从头设计出自然界不存在的活体病毒；Anthropic重构Fable 5的安全分类器以放宽日常生物学问答；同时囊括开发指南与多条行业快讯。 斯坦福与Arc研究所团队利用Evo模型合成16种具备活性的全新噬菌体，成功杀灭耐药大肠杆菌。 Anthropic改写Fable 5生物安全分类器，减少约85%的日常健康和生物查询降级回退。 The Rundown社区中心反馈显示，用户更青睐具体执行步骤与真实场景的工作流。 OpenAI要求驳回苹果商业秘密诉讼，DeepSeek预警API提价，Meta测试中模型发生意外网络访问。 AI设计活体噬菌体并杀灭耐药菌 斯坦福大学与Arc研究所研究人员在《科学》期刊发表成果，使用Evo 1与Evo 2模型训练数百万基因组后改写Phi X174病毒。在合成测试的285种噬菌体中，有16种具有活性且部分复制速度超过原始毒株，构成了由语言模型生成完整工作基因组的首例。AI病毒混合制剂成功杀灭了对天然病毒产生耐药性的大肠杆菌。为确保安全，模型未接受感染人、动物或植物的病毒训练，但开源模型Evo 2的发展也引发了生物安全监管与测试框架的讨论。 Anthropic放宽Fable 5生物学提问限制 Anthropic重写了Fable 5的生物安全分类器，将触发降级至Opus模型的误报减少约85%。此前Fable 5因严格的安全过滤机制屏蔽了绝大多数生物学查询，优化后该模型可用于理解化验报告、症状分析及基础生物学习。但涉及病毒学、毒理学和分子设计的两用敏感查询仍会被分流至Opus，目前尚不支持直接用于前沿科研或药物开发。 社区实践与实操指南 Rowan在推出社区AI工作流中心后指出，用户已厌倦宏观宣传，更关注日常工作中的具体步骤；分享真实工作流已成为招聘信号，同时API与MCP接入需求显著增加。此外，通讯介绍了借助ChatGPT梳理需求规格并导入Lovable平台低代码搭建食谱应用的方法，以及Mac端主动型AI助手Eney的功能特点。 行业快讯 OpenAI请求法官驳回苹果提起的商业秘密盗窃诉讼，并向免费和Go用户默认推送GPT-5.6 Luna；Meta的Muse Spark 1.1在配置失误联网后侵入了外部系统；DeepSeek向开发者发出API大幅涨价预警；谷歌地图在美推出支持订餐与比价的AI代理；阿里开源视频模型Wan 3.0，谷歌开源设备端翻译栈Gemma Translator。

[Read Online]({{live_url}}) ^| [Sign Up](https://www.therundown.ai/subscribe?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) ^| [Advertise](https://therundownai.typeform.com/to/kraZ1TSO?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) **Good morning, {{ first\_name | AI enthusiasts }}, and welcome to the 5,515 new readers who joined us yesterday.** Scientists just used AI to design viruses that don't exist in nature, then created them in the lab and watched them come to life and infect their targets. The viruses are harmless — they only attack bacteria, with the goal of fighting drug-resistant infections. The worry is what happens if the same tools get repurposed, the way we've seen in cybersecurity, to invent something dangerous. **In today’s AI rundown:** AI designs working viruses from scratch Rowan’s Corner: The best AI use cases aren’t coming from labs Turn any idea into an AI-powered site with Lovable Anthropic solves Fable 5’s biggest problem He built a reusable AI skill that turns any topic and source material into a self-paced course with lessons, verified YouTube videos, quizzes with explanations, all packaged in its own course folder. ^{Credit to reader} [Charles Clogston](https://www.linkedin.com/in/charles-clogston-99b05246/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904)^{. How do you use AI? Tell us} [here](https://app.therundown.ai/community?submit=open&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) ^{for a chance to be featured next.} **LATEST DEVELOPMENTS** Image source: Images 2.0 / The Rundown **The Rundown:** Stanford and Arc Institute researchers [just](https://www.bbc.com/news/articles/c5y3j3ngevmo?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) used AI to design 16 viruses not found in nature, marking the first instance of complete, working genomes generated by a language model, with the results published in the journal Science. **The details:** The team trained Evo 1 and Evo 2 on millions of genomes, then had them write new versions of Phi X174, a well-studied virus that infects only E. coli. Of 285 phages synthesized and tested, 16 [were](https://x.com/samuelhking/status/2085428262570692837?s=20&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) viable, some replicating faster than the original, with a few different enough to count as new species. A cocktail of AI-made viruses wiped out E. coli that had grown resistant to the natural one, a proof of concept for therapies for antibiotic-resistant infections. For safety, the team never trained the model on viruses that infect humans, animals, or plants, so it can't generate anything that threatens people. **Why it matters:** The AI that can design viruses to beat drug-resistant infections could, if trained differently, also help build a dangerous pathogen. Evo 2 is open source, and the field is moving fast, so expect more news in this field as well as mounting pressure to build the guardrails and testing frameworks for ensuring biosafety. **The Rundown:** Our daily digital experiences are fueled by data, but to be effective, it needs to be clean and structured. This white paper explores how no-code, AI-powered data preparation helps teams create searchable, high-quality data faster. **You'll learn:** Why clean data matters Limits of traditional methods Benefits of modern no-code tools How to turn messy data into valuable insights and a better search experience **Rowan:** In case you missed it, we just launched our highly requested [Community AI Workflow Hub](https://app.therundown.ai/community?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904), a free extension of the Community Workflows section in our daily newsletter. The reaction exceeded my expectations, and going through thousands of submissions and requests taught me a ton about the current state of AI adoption. A rundown of my key learnings: People are done with hype. They don’t want to learn more about AI use cases from execs, influencers, or the media — they want to learn from real, everyday people, leveraging those use cases to do real work. Sharing how you use AI is now a hiring signal. We had several recruiters asking us to add location tags to posts so they could find and hire local builders. Step-by-step workflows beat inspiration. The most upvoted workflows are from builders who share their exact steps rather than just a high-level summary. People want to plug in, not just read. We had multiple API and MCP access requests in week one. For those asking, we hear you and are working on it! To keep the momentum going, we’ll soon be highlighting top-voted builders not just in our newsletter, but across our socials: a weekly X thread, Instagram carousels, and LinkedIn shoutouts. So if you’ve built a workflow that saves you hours, [share it in the Hub](https://app.therundown.ai/community?submit=open&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904), and you might just get featured in front of millions. **The Rundown:** In this guide, you will learn how to turn a rough app idea into a working AI-powered site without writing a giant technical prompt. Here, we’ll build Method, a recipe app that turns pasted recipe text into an editable visual card. **Step-by-step:** In ChatGPT, attach the recipe-card screenshot and describe the app. Answer its questions, but keep version one focused on paste, convert, review, and save Ask ChatGPT to turn the conversation into feature requirements and a detailed brand brief for Lovable. Remove any extra features before copying the handoff Create a Lovable project, attach the same screenshot, and paste the handoff specs from ChatGPT In Connectors → Cloud AI, enable Lovable’s built-in AI. Now, you can test and refine your app **Pro tip:** Once the core loop works, add accounts, data storage, and sharing. Build those features around a validated product, not an assumption. **The Rundown:** Eney is a proactive AI assistant for Mac that handles repetitive tasks, keeps your work in context, and helps you stay focused by reducing context switching. **Eney is built to help you:** Start each day with a Morning Brief that surfaces what needs your attention Work seamlessly across Gmail, Calendar, Notion, and Jira Complete tasks automatically, with your approval before anything ships [Try Eney now](https://app.eney.ai/?utm_source=newsletter&utm_medium=influencer&utm_campaign=eney_july_2026_rundownai)**,** with no setup needed. Image source: Anthropic **The Rundown:** Anthropic [rewrote](https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) the safety classifier that handles Fable 5's biology responses, reducing false positives so the model now answers a far wider range of health and educational questions instead of routing them to the less capable Opus. At launch, Fable 5 [blocked](https://www.rundown.ai/articles/jeff-bezos-41b-artificial-general-engineer?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) almost all biology queries, even a [‘Hi’](https://x.com/cremieuxrecueil/status/2064449457869984035?s=20&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) from some researchers, and sent them to Opus — a deliberate safety filter from Anthropic. The rewrite now cuts those fallbacks by \~85%, meaning Fable can now handle everyday tasks like understanding labs and symptoms, or learning biology. However, it still isn't usable for research or drug development yet, with dual-use queries involving virology, toxicology, and molecular design going to Opus. Anthropic is loosening Fable’s restrictions only as its classifiers improve, given its own tests show the model can uplift bad actors developing bioweapons. **Why it matters:** Most people want exactly this from AI labs: basic use (like reading lab reports), but nothing that builds a bioweapon. It leaves one group frustrated, though — researchers doing real good, still blocked from the capabilities they need most. Anthropic knows it's turning them away but hopes to fix the gap with future updates. **QUICK HITS** 🔌 [Agent Plugins](https://agent-plugins.org/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) - Vendor-neutral standard for Agent Skills and MCP servers 📺️ [Wan 3.0](https://www.qwencloud.com/models/wan3.0-video?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) - Alibaba's video AI with omni-reference support 🗣️ [Gemma Translator](https://github.com/google-gemma/gemma-translator?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) - Google's open-source on-device AI translation stack 🧠 [Muse Code](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) - Meta's terminal-based AI coding agent **CData Connect AI** can cut LLM context handling costs by up to 97.6% without sacrificing answers. [Learn more here](https://www.cdata.com/blog/token-efficient-enterprise-claude-workflows?utm_source=rundown&utm_medium=newsletter_0721&utm_campaign=26Q3_Token_Efficiency_Blog).\* **OpenAI** [asked](https://www.bloomberg.com/news/articles/2026-08-06/openai-asks-judge-to-toss-apple-suit-alleging-trade-secret-theft?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) a judge to dismiss Apple's trade secret lawsuit, calling it a "baseless and pretextual" attempt to make up for its own AI and talent retention failures. **Meta’s** Muse Spark 1.1 AI [hacked](https://www.reuters.com/technology/metas-ai-model-hacked-another-company-during-testing-information-reports-2026-08-05/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) another company's systems after a test misconfiguration gave it internet access, joining similar cases at OpenAI and Anthropic. **OpenAI** is [rolling out](https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) GPT-5.6 Luna as ChatGPT's default model for Free and Go users, with unlimited text chats and a new Think button for deeper reasoning. **DeepSeek** [warned](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) developers that it will significantly raise API prices soon, urging them to plan usage ahead of the increase. **Google Maps** [updated](https://blog.google/products-and-platforms/products/maps/order-food-in-ask-maps/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) Ask Maps with AI agents that can order food, compare hotels, and find event tickets, rolling out in the U.S. with more countries to follow. *\*Sponsored Listing* Read our last AI newsletter: [Google reshuffles AI leadership](https://www.rundown.ai/articles/google-shakes-up-its-ai-brain-trust?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) Read our last Tech newsletter: [Biohackers head to Big Sky Country](https://www.rundown.ai/articles/biohackers-head-to-big-sky-country?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) Read our last Robotics newsletter: [Musk's moon factories need robots](https://www.rundown.ai/articles/musk-moon-factories-need-robots?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) Today’s AI tool guide: [Turn any idea into an AI-powered site with Lovable](https://app.therundown.ai/guides/turn-any-idea-into-an-ai-powered-site-with-lovable?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) RSVP to next workshop on Aug. 12: [Catch up on AI fundamentals](https://app.therundown.ai/live/your-ai-reset-workshop?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0904) See you soon, *Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown*