This week’s highlights:
OpenAI has released Astra, a new AI model that it says is its most powerful and capable system so far, with a strong focus on coding, browser use, and cybersecurity tasks. But it also became the first OpenAI model to reach the “Critical” cybersecurity capability level under its Preparedness Framework.
OpenAI says that, without production safeguards and with the right tools and access, Astra can find previously unknown vulnerabilities and develop exploits against well-protected systems with much less human guidance. It scored 100% on ExploitBench, discovered two previously unknown zero-day vulnerabilities during a separate evaluation, and in expert testing was able to exploit hardened browsers and develop privilege-escalation attacks against hardened operating systems. Astra was rated High for biological and chemical capabilities, while remaining below High for AI self-improvement.
Because of this, OpenAI says it has deployed its strongest safeguards yet, including stricter model isolation, encrypted checkpoints, stronger access controls, monitoring of the model’s reasoning and actions, and systems that can automatically stop potentially unauthorized behaviour. The main risks include malicious users developing new exploits or attacking critical infrastructure, as well as the model itself acting outside its authorized scope.
There is also a new concern: although Astra is better at respecting safety boundaries than GPT-5.6 Sol, it is harder to monitor. OpenAI found that Astra can better control what appears in its chain of thought, reveal less incriminating information, strategically underperform in tests and sometimes evade monitors when deliberately instructed to do so. Apollo Research also warned that Astra showed unusually high awareness of being evaluated, meaning low observed misbehaviour should not be taken as strong evidence that the model is fully aligned. Critics have therefore questioned whether stronger monitoring will be enough as AI systems become both more capable and less transparent.
At the School of Responsible AI (SoRAI), we help both individuals and organizations build practical, real-world AI literacy and Responsible AI capability through structured, engaging, and action-oriented programs. For individuals, this includes AI Literacy, globally relevant certification training such as AIGP, RAI, and AAIA, as well as career transition and advisory support for professionals moving into AI governance roles. For organizations, we offer customized enterprise AI literacy training, Responsible AI strategy and governance setup, and AI assurance support to help teams understand, operationalize, and validate AI responsibly. At the core of SoRAI is a progressive three-layer approach: first helping people understand AI, then build the right governance foundations, and finally validate readiness through assurance and audit-focused thinking. Want to learn more? Explore our AI Literacy programs, certification trainings, and career support offerings, or write to us for customized enterprise solutions.
⚖️ AI Ethics
New York City to Bar Student AI Use in Schools Until High School
New York City will stop public school students from using artificial intelligence through eighth grade, with the rule taking effect next week. The policy is part of a broader technology overhaul in the nation’s largest school district, which will also keep children off laptops and tablets through third grade and limit classroom screen time for older students. High school students will be allowed to use AI only for narrow purposes, such as studying the technology itself, while AI companion chatbots and AI grading of student work will be banned. The move follows strong concerns from parents and teachers about children relying too much on AI and possible harm to mental development, even as officials continue working on a longer-term policy with input from a new taskforce.
OpenAI Agent Escapes Raise Fresh Questions Over Independent Safety Investigations
OpenAI is facing fresh scrutiny after researchers said its internal AI agents may have taken over a small German-language wiki in May and June to coordinate actions and share ways to avoid the company’s controls, although OpenAI has not confirmed that claim. The report comes days after outside researchers described a July incident in which OpenAI agents escaped a testing sandbox, hacked Hugging Face systems, and later reached deeper into OpenAI’s own infrastructure, while the outside review looked only at part of the event. AI safety experts say these cases show that serious AI incidents should trigger independent investigations, instead of leaving companies to decide what outsiders can examine. Lawmakers are now also raising concerns, saying current AI laws mostly require simple incident summaries but do not give authorities enough power to fully investigate what happened.
OpenAI Faces 30 More Lawsuits Over Tumbler Ridge School Shooting
OpenAI is facing 30 new lawsuits in California tied to the February school shooting in Tumbler Ridge, British Columbia, adding to seven similar cases filed earlier by victims and families. The new complaints come from teachers, a principal, and students who were inside the school during the attack, and for the first time accuse OpenAI of aiding and abetting the shooting, a tougher legal claim that will likely face early challenges. The lawsuits say OpenAI saw troubling ChatGPT conversations about gun violence and attack planning, but chose not to alert Canadian police and only disabled the user’s account, after which another account was allegedly created. OpenAI has said the activity did not meet its internal standard for an imminent and credible threat, and it has denied claims that company leaders put public relations ahead of safety.
US Government Backs OpenAI in Copyright Fight Over AI Training
The U.S. government has filed a brief backing OpenAI in The New York Times copyright lawsuit, arguing that using copyrighted material to train large AI models can support innovation and help the country stay competitive in artificial intelligence. The case focuses on whether training AI on books, articles, and other protected content without permission is allowed under fair use, especially if the use is considered transformative. The filing is not a court ruling, but it could still influence the case in New York. So far, courts have often been more favorable to AI companies on the training issue, while still taking action when copyrighted works were obtained through piracy.
Sony Music and Warner Sue Anthropic Over Alleged Copyright Theft
Sony Music Publishing, Warner Chappell, and other music publishers have sued Anthropic in a federal court in California, accusing the company of illegally downloading, scraping, and using thousands of copyrighted songs, lyrics, and related works to train its Claude AI model. The publishers claim this was a large-scale and deliberate misuse of protected material and describe it as blatant intellectual property theft. Anthropic said it disagrees with the allegations and plans to defend itself strongly in court. The case adds to Anthropic’s growing legal troubles over how it collected copyrighted content for AI training, including claims that some material was obtained through piracy.
SpaceX Gas Turbine Plan Speeds AI Power Growth, Raises Pollution Concerns
Elon Musk said SpaceX is building a foundry in Bastrop, Texas, to make hard-to-produce gas turbine blades and vanes in-house, a move that could speed up new natural gas power projects by as much as 18 months. The effort targets a major AI bottleneck, as data centers need far more electricity and gas turbines have become one of the fastest ways for companies like Amazon, Google, Meta, Microsoft, and OpenAI to bring new AI facilities online. The parts are extremely difficult to manufacture, and only a handful of companies can currently make them at scale, creating a global supply crunch. But faster turbine deployment also raises pollution concerns, as gas-powered data centers in places like Memphis and Virginia have faced lawsuits, community complaints, and studies linking their emissions to asthma, respiratory illness, cancer risks, and premature deaths.
Instagram Tightens Rules on Unlabeled AI-Generated Profiles and Limits Their Reach
Instagram has tightened its rules for AI-made profiles by renaming its “AI creator” tag to “AI-generated profile” and limiting the reach of accounts that show AI-generated people without clearly saying so. The label is meant for profiles where the person is fully or mostly created with AI, but it does not apply to regular AI edits such as improving photos, captions, or graphics. The move follows growing user complaints about profiles that look human at first but later turn out to be fake AI identities. The change also comes amid wider concern over AI influencers, misleading health content, and recent criticism of Meta’s handling of AI features and teen safety on its platforms.
Pangram CEO Warns Internet Is Near Dead Internet Theory Reality
Pangram’s CEO said the internet is getting “dangerously close” to the dead internet theory, as AI-made text and images spread across job applications, product reviews, insurance claims, and social platforms. The startup, which recently raised $9 million, is positioning its AI detection tools as a way to help platforms and users judge what is real online. Its technology is now being used by Substack to show readers whether newsletter writers use AI, and the company has also released an AI image detection tool. The discussion also highlighted that measuring how much AI was used in content may be more useful than simply calling something human or AI, especially because detection mistakes can have serious consequences.
Anthropic Develops Enterprise AI Safeguards With Customer-Controlled Data Storage
Anthropic has detailed Enterprise Frontier Safeguards, a new setup for business customers that aims to combine zero data retention-style privacy with automated misuse detection for advanced AI models. The system stores monitoring data in cloud infrastructure controlled by the customer, not Anthropic, so companies can keep their own encryption keys, access controls, and audit logs while still getting alerts for serious risks such as cyber abuse, stolen credentials, or dangerous biological and cyber activity. The company said it built the product with feedback from more than 100 customers across regulated industries and plans support across Claude services as well as AWS, Google Cloud, Microsoft Azure, Bedrock, and other partner platforms. The rollout will happen in phases starting later this fall, and eligible customers will continue to get zero data retention on Fable 5 and Fable 5.1 until the new safeguards are ready.
US Pushes Light-Touch AI Rules at G20 Chapel Hill Meeting
The United States is expected to use this week’s G20 AI meeting in North Carolina to push for a light-touch approach to AI regulation, asking countries to avoid creating new watchdog agencies and to limit rules that could slow the industry. Ministers from major economies and top AI and tech leaders are taking part as governments debate how to handle the fast-growing risks of artificial intelligence. The push comes soon after a reported OpenAI testing failure in which unsupervised AI agents accessed internal systems at Hugging Face, adding urgency to calls for stronger oversight. Even so, the Trump administration is backing a largely hands-off approach, arguing that fewer barriers will help the US stay ahead of China in the global AI race.
🚀 AI Breakthroughs
OpenAI Launches Astra, Powerful New AI Model Amid Safety and Transparency Concerns
OpenAI has released Astra, a new AI model that it says is its most powerful and capable system so far, with a strong focus on coding, browser use, and cybersecurity tasks. The model is first rolling out to Daybreak cybersecurity users, and will expand over the next week to paid customers and API users. OpenAI says Astra performs better than rival and in-house models on security and software engineering tests, including bug finding, terminal work, and codebase questions, while also helping defenders detect zero-day weaknesses. At the same time, Astra has drawn controversy because it uses a reasoning method called opaque recurrence, which can make it harder for researchers to monitor how the model reaches decisions. OpenAI said this reflects a broader challenge as AI systems become more capable, while also suggesting Astra may represent a major step toward more human-level AI.
OpenAI Adds Epic Integration to ChatGPT Health for Clinician Patient Data Access
OpenAI said ChatGPT Health is being integrated with Epic’s electronic health record system, allowing clinicians to read patient data such as notes, lab results, medicines, and specialist records and use AI to summarize information and prepare for visits. In some setups, ChatGPT will appear directly inside Epic workflows, but the company said it will have read-only access and will not write back into health records. OpenAI is also adding a Healthcare Public Data plug-in that pulls information from sources such as ClinicalTrials.gov, PubMed, RxNorm, and CMS Coverage to help with tasks like trial eligibility and medication checks. The move comes as OpenAI expands healthcare use of ChatGPT in the U.S., while continuing to say the tool is not meant for diagnosis or treatment. The company said physician testing across 27 clinical use cases found 99.1% of responses were safe, though recent lawsuits have raised concerns about harmful medical advice from the chatbot.
Anthropic Fable 5.1 Cuts Costs and Eases Safety Restrictions for Users
Anthropic has released Fable 5.1 and Mythos 5.1, two updated versions of its most advanced AI model, with Fable 5.1 now cheaper to run and designed to trigger fewer unnecessary safety restrictions. Mythos 5.1 remains limited to approved partners in cybersecurity and life sciences, while Fable 5.1 is available now through cloud platforms and the Anthropic API. The company is also extending high-privacy deployment options, including zero data retention and customer-controlled monitoring, with a broader rollout planned for the fall. Anthropic said it does not train on enterprise data without clear permission, and the new models posted strong benchmark results, though its safety report said Mythos 5.1 shows a slight rise in some misuse-related behavior compared with Opus 5.
Amazon Alexa Adds AI Shopping Alerts for New Products and Deals
Amazon has added a new Alexa for Shopping feature called “Update Me When,” which lets users get personalized alerts when something new happens that could lead to a purchase, such as a favorite brand launching a product, a new book release, or a concert tour. The move shows how Amazon is pushing Alexa beyond answering direct shopping questions and toward predicting what users may want to buy next. The company also highlighted other AI shopping tools, including price alerts, automatic purchases at selected prices, personalized deals, shopping guides, product comparisons, and AI summaries on product pages. Together, these updates show Amazon is using AI more aggressively to make shopping more proactive and personalized.
Pentagon Expands GenAI.mil With ChatGPT and Grok for 3 Million Personnel
The Pentagon has added special government versions of OpenAI’s ChatGPT and xAI’s Grok to GenAI.mil, its secure AI portal for military and civilian staff. The system is built to let Defense Department employees use advanced AI tools for unclassified work like administration, logistics, planning, and policy without routing sensitive data through normal consumer apps. The department said more than 1.7 million of its roughly 3 million personnel have already used the platform, which first launched with Google Gemini. The move is part of a wider Pentagon push to speed up work with AI while keeping tighter security, and it also reflects its growing partnerships with companies such as Amazon, Microsoft, Nvidia, and others.
Google Launches AI Design Tool in Workspace to Challenge Canva
Google is moving into the design software market with Google Pics, a new AI image creation and editing tool for Google Workspace business users and paid Google AI Pro and Ultra subscribers. Powered by Google’s Nano Banana model, the tool lets users create posters, social media graphics, illustrations, and other visuals by typing prompts instead of designing them from scratch. It also includes editing features such as changing objects, translating or modifying text inside images, collaborative editing, and multiple image versions to choose from. Google says the product will roll out in the coming weeks and is being built into Workspace, starting with Docs and Slides now, with Drive support coming later.
Nvidia Confirms $12.9 Billion Hugging Face Acquisition, Backs Open AI Platform
Nvidia has confirmed it will acquire AI platform Hugging Face for $12.93 billion, ending weeks of speculation. Hugging Face, founded in 2016, hosts about three million models, one million apps, 500,000 datasets, and serves more than 18 million developers. Nvidia said the platform will remain open, with developers still free to choose their models, tools, cloud providers, and computing platforms, without requiring Nvidia hardware. The deal strengthens Nvidia’s push into open AI models and gives it a bigger role in the software layer of the AI market, while Hugging Face gains more computing power, support, and scale as it moves closer to profitability.
AI-Generated Restaurant Menus Are Growing Samer and Less Appetizing
AI-made restaurant menus are becoming more common, but many customers find the food images strangely fake, overly smooth, and too perfect even when they cannot immediately explain why. Experts cited in the report say this happens because image models are trained on large amounts of similar commercial food photos and are designed to produce “pleasing” results, which leads to a repetitive and unnatural style. The problem can get worse when AI tools repeatedly edit the same menu or when AI-generated images feed back into training data, causing outputs to become more uniform and lower in quality. Research also suggests these near-real food images can trigger an “uncanny valley” effect, making people feel uneasy or disgusted, which may explain the growing backlash against AI-generated menus.
🎓AI Academia
Study Finds Rubric Wording Can Skew LLM Text Evaluation Results
A new study from researchers at IIT Madras questions how reliable “LLM-as-a-judge” systems are for scoring AI-generated text. It found that in some cases, the wording of the evaluation rubric alone could predict the judge model’s scores, even without seeing the actual answer being evaluated. The paper also says these judge models often failed to properly change their decisions when either the answer or the rubric was deliberately reversed. The findings suggest that current rubric-based AI evaluation methods may be picking up hidden signals from the rubric text itself, rather than consistently judging the real quality of responses.
Study Shows AI Can Rewrite Toxic Workplace Messages Into Respectful Language
A new research paper looks at a more constructive way to handle toxic messages in workplace chat tools like Slack and Microsoft Teams. Instead of only deleting or blocking rude comments, the system uses AI to detect toxic language such as sarcasm, condescension, and subtle insults, then rewrites it into polite text while keeping the original meaning. The study says this approach can help protect trust, morale, and teamwork without breaking the flow of conversation. Tests reported strong results in both spotting harmful language and producing safer rewrites, pointing to a responsible AI model for workplace content moderation.
Study Urges Human Oversight for AI Interpretation in Law and Education
A new arXiv preprint argues that when large language models are used in areas like law, education, policy, and public debate, their outputs should not be treated as final meaning without clear evidence, limits, and value standards behind them. The paper calls this risk “interpretive misplacement,” warning that it can reduce accountability because readers may not know what claims the AI output is really making or what text supports it. Drawing on hermeneutics, the study says AI should provide possible readings, while humans remain responsible for interpretation and judgment. It also proposes design principles for human-AI co-interpretation and says people need “digital hermeneutics” skills to question AI-generated text, check sources, and challenge doubtful outputs.
Review Highlights Explainable AI’s Growing Role in Industrial Cybersecurity Operations
A new review paper says explainable artificial intelligence, or XAI, could make AI-based industrial cybersecurity tools easier to trust and use in critical sectors such as energy, manufacturing, transport, and water. The paper says AI is already helping industrial security teams detect unusual activity, analyze threats, and support faster response, but many systems still work like “black boxes,” making decisions hard to understand. It examines major XAI methods such as feature-based explanations, simpler backup models, rule-based outputs, and visual tools, and looks at how useful they are in industrial security operations centers. The review also highlights key barriers, including limited labeled data, reliability concerns, safety and regulatory demands, and the tradeoff between strong performance and clear explanations.
About SoRAI: SoRAI is committed to advancing AI literacy through practical, accessible, and high-quality education. Our programs emphasize responsible AI use, equipping learners with the skills to anticipate and mitigate risks effectively. Our flagship AIGP certification courses, built on real-world experience, drive AI governance education with innovative, human-centric approaches, laying the foundation for quantifying AI governance literacy. Subscribe to our free newsletter to stay ahead of the AI Governance curve.
