Trump signs executive order that invites vetting of top AI models
++ MIT reports warn catastrophic AI threats remain likely; Florida sues OpenAI and Sam Altman; Amazon Ring faces facial-recognition privacy lawsuit & more..
Today’s highlights:
President Donald Trump has signed a narrower executive order on AI oversight after objections from industry leaders, asking certain AI companies to voluntarily submit powerful new models to the government for testing 30 days before public release. The final order is less strict than an earlier draft that proposed a 90-day review window, while also making clear that it does not create any mandatory licensing or pre-approval system for AI models. The move follows a delay in late May as the administration reconsidered the tougher version amid concerns it could slow US AI firms in competing with China. The order also tells the Justice Department to prioritize crimes involving AI-assisted hacking and unauthorized access, adding to the administration’s broader push to shape a national AI policy framework.
Three days after issuing its executive order, the White House also released National Security Presidential Memorandum 11 to speed up AI adoption across the US military and intelligence agencies. The memo says agencies should cut barriers to deployment, use advanced models from multiple vendors, expand secure computing infrastructure, and build partnerships with private companies to protect sensitive AI systems. It also orders updates to weapons autonomy policy, procurement rules, testing standards, and workforce training, while stressing that AI use must remain reliable, controllable, lawful, and aligned with civil liberties. The directive replaces earlier guidance and sets deadlines of 90 to 120 days for agencies to deliver new policies, roadmaps, and security measures.
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
Anthropic Says AI Is Accelerating Its Own Development, Raising Recursive Self-Improvement Concerns
Anthropic says AI is increasingly helping build newer AI systems, speeding up software engineering and some research tasks inside the company. In a new blog post, it claims engineers now ship far more code with AI assistance, with Claude reportedly authoring more than 80% of merged code by May 2026, while benchmark results also show models handling longer and more complex coding and research tasks. The post argues this trend could eventually lead to “recursive self-improvement,” where an AI system can design its own successor, though it says that point has not yet been reached and may not be inevitable. Anthropic also warns that if such systems arrive, they could bring major gains in science and healthcare but also raise serious risks around oversight, control, and the need for global coordination or possible slowdowns in frontier AI development.
MIT AI Risk Initiative Report Finds Experts Warn of Severe AI Threats Across Key Sectors
MIT’s latest AI Risk Initiative report, based on a Delphi survey of 272 international experts, finds that many AI risks could become severe within five years, with 18 of 24 risk areas carrying at least a 10% chance of catastrophic outcomes under current trends. Experts ranked dangerous capabilities, competitive dynamics, weapons and cyberattacks, power centralization, and false information among the most serious threats, while saying practical mitigations would reduce harm but not eliminate major risks. The report says the public and AI users are the most exposed to harm, but responsibility for fixing these problems falls mainly on general-purpose AI developers and governments, creating what it describes as a responsibility gap. It also identifies information, finance, and national security as the sectors most vulnerable to AI-related disruption, and argues that stronger rules, enforcement, and international coordination will be needed to manage the risks.
Florida Sues OpenAI and Sam Altman Over ChatGPT’s Alleged Links to Violent Incidents
Florida has filed what it describes as the first state-led lawsuit against OpenAI and CEO Sam Altman, alleging ChatGPT was released despite safety warnings and has been linked to several violent and harmful incidents. The complaint says the company misled the public about the chatbot’s risks and allowed a dangerous product to reach millions, including children, while prioritizing growth and profits. The case follows a criminal probe opened in April into ChatGPT’s alleged role in the 2024 Florida State University mass shooting, where the suspect is said to have consulted the bot before the attack. OpenAI has denied responsibility for that shooting, while facing other lawsuits that similarly seek to connect ChatGPT to suicides, stalking, and murder.
Amazon Faces Class Action Lawsuit Over Ring Facial-Recognition Feature and Alleged Privacy Violations
Amazon is facing a proposed class action lawsuit that alleges its Ring doorbells violate privacy laws through the Familiar Faces feature, which uses facial recognition to identify frequent visitors and store images of people who may not have consented. The lawsuit, filed in Seattle, argues that millions of people may have had facial data collected simply by passing Ring cameras. Ring launched the opt-in feature in December despite criticism from privacy advocates and lawmakers, while saying face data is encrypted, not shared, and unknown faces are deleted after 30 days. The case adds to Ring’s broader privacy scrutiny, following a 2023 FTC settlement in which Amazon paid $5.8 million over claims that employees and contractors improperly accessed customers’ private videos.
Google Rolls Out Android Fake Call Detection to Counter AI Deepfake Impersonation Scams Globally
Google has started rolling out a fake call detection feature in Phone by Google on Android 12 and newer devices, beginning with Pixel phones, to help protect users from AI deepfake impersonation scams. The tool is enabled by default and uses a silent verification process between devices to confirm whether a call from a saved contact is genuinely coming from that person’s phone. If the signal is missing and the real device shows no call is being made, Android warns the user that the call may be fraudulent. Google said the system is built on RCS, which could allow broader adoption, and the update comes alongside new Android features for Google Photos, Play Books, and Circle to Search.
Martin Scorsese Joins Black Forest Labs as Adviser, Uses AI Only for Film Storyboarding
Martin Scorsese has joined German AI image startup Black Forest Labs as a partner and adviser, but his use of the technology is limited to creating storyboards for films, according to The New York Times. Scorsese said the tool helps him communicate visual ideas to cinematographers and production designers more quickly, while noting he has long made his own storyboards. Black Forest Labs, founded by the team behind Stable Diffusion, powers image tools for companies including Adobe, Canva, Microsoft, and Meta, and was recently valued at $3.25 billion. The move is likely to raise concerns in parts of Hollywood, but it also reflects a broader easing of the entertainment industry’s earlier resistance to AI.
Uber Imposes Employee AI Spending Caps After Exhausting Annual Budget in Just Four Months
Uber has introduced a new internal limit on employee use of AI coding tools, capping spending at $1,500 per month per employee for tools such as Claude Code and Cursor, according to Bloomberg. The move comes after the company reportedly used up its full annual AI budget in just four months after encouraging staff to use AI heavily and tracking usage through internal leaderboards. Employees can monitor their usage through an internal dashboard, and exceptions to the cap may be granted in some cases. The decision highlights growing pressure on large companies to control rising AI costs as questions remain over whether heavy enterprise AI spending is delivering clear returns.
UK Regulation Requires Google to Let Publishers Opt Out of AI Search Results
Google said it will let publishers opt out of appearing in its generative AI Search features in the U.K., following new requirements from the U.K. Competition and Markets Authority. The opt-out will be managed through a new Search Console setting, and sites that use it will be excluded from features such as AI Overviews, AI Mode, and AI Overviews in Discover, without affecting their rankings in regular Google Search. The regulator said the move is a first-of-its-kind step aimed at giving publishers, including news organizations, more control over how their content is used and more leverage in content negotiations. Google also said it will improve attribution in AI results with clearer links and provide publishers with new Search Console metrics showing how their pages appear in AI responses, with an initial U.K. test planned before a broader global rollout.
Supreme Court Draft AI Rules Permit Court Use, Bar AI From Judicial Decision-Making
The Supreme Court has issued draft 2026 rules allowing artificial intelligence tools in courts for tasks such as legal research, drafting, translation, transcription and case management, while clearly banning their use in deciding cases or predicting judicial outcomes. The framework says judges alone must determine questions of law, facts and justice, and that AI can only play an assistive role under human oversight. It also proposes that lawyers disclose when they use AI in preparing pleadings or evidence, and restricts opaque AI systems in high-risk matters affecting personal liberty or legal rights. The draft, prepared by a Supreme Court committee, has been opened for public consultation, with comments invited until June 20.
India Plans AI-Driven Machine-Readable SMART Standards to Reduce Industry Compliance Burden, Consumer Affairs Ministry Says
India’s Consumer Affairs Ministry plans to use artificial intelligence to create machine-readable SMART standards that can turn regulations into digital rules for faster and more accurate compliance checks. Speaking at an industry event in New Delhi, a senior official said the move is aimed at reducing the compliance burden on businesses and modernising India’s standards system. The government said the Bureau of Indian Standards is shifting from a regulator-led role toward a more facilitative approach, while also encouraging private sector testing infrastructure. It is also working to replace outdated testing methods with quicker and more reliable processes as part of its broader quality and industrial growth goals.
Teradata Freezes 2026 Salary Hikes for 5,100 Employees as Funds Shift to AI Investment
Teradata has told about 5,100 employees they will not receive annual salary hikes in 2026, with the company instead redirecting that budget to expand its artificial intelligence talent, products and technology, according to Business Insider. Workers had typically seen yearly raises of around 2% to 4%, though such increases were not guaranteed, and the salary freeze will mainly apply in regions where employers are not legally required to make market-based adjustments. The company said employees may still receive performance bonuses, equity awards and other incentives even as base pay remains unchanged. The move reflects a wider trend across the tech industry, where companies facing weaker revenue growth and economic pressure are prioritising AI investment as a key driver of future competitiveness.
Global Media Coalition Expands Across Europe and North America to Demand Fair AI Payment
About 30 media outlets from Europe and North America have joined the SPUR Coalition, a group led by major publishers including the BBC, Sky News and The Guardian to push for fair payment from AI companies for the use of news content. The coalition says AI firms use costly publisher content to train large language models without proper permission or compensation, raising fresh concerns over the future of journalism’s business model. SPUR is seeking stronger content protection, fair value-sharing, and systems that let publishers track how their material is used by AI tools. It also plans to work on licensing frameworks that could give AI developers a clearer and more sustainable way to pay for news content.
🚀 AI Breakthroughs
OpenAI Expands Codex With Enterprise Plug-Ins and Workplace Tools Beyond Software Engineering
OpenAI has expanded Codex with six new job-focused plug-ins for data analytics, creative work, sales, product design, equity investing, and investment banking, as it steps up efforts to attract enterprise users beyond software developers. The company said Codex now has more than 5 million weekly active users, with knowledge workers making up about 20% of users and growing more than three times faster than developers. OpenAI also added a Sites feature that lets Codex turn work into hosted interactive websites, along with Annotations for more precise editing and commands inside documents and files. The move follows recent enterprise-focused efforts by OpenAI, including a new deployment venture backed by more than $4 billion, as competition in workplace AI tools intensifies.
OpenAI Launches Lockdown Mode to Reduce Prompt Injection Risks for Sensitive ChatGPT Data
OpenAI has rolled out Lockdown Mode, a new ChatGPT security feature aimed at reducing the risk of prompt injection attacks that can hide malicious instructions in webpages and other content sources. When enabled, the mode turns off live web browsing, web image retrieval and display, deep research, and agent mode, while still allowing access to cached content and image generation. OpenAI said the feature does not fully eliminate prompt injection risks, since harmful instructions could still appear in cached web pages or uploaded files and influence responses. The company said Lockdown Mode is meant for users and organizations handling sensitive data and is currently being rolled out to self-serve ChatGPT Business accounts and eligible personal accounts.
ChatGPT App Reaches 1 Billion Monthly Users at Record Pace, Sensor Tower Data Shows
OpenAI’s ChatGPT app crossed 1 billion global monthly active users in May, about three years after launch, making it the fastest app to reach that mark, according to Sensor Tower estimates. The milestone comes as competition in generative AI intensifies, with Anthropic’s Claude also expanding quickly. Sensor Tower said U.S. ChatGPT users who installed Claude in the first quarter of 2026 spent 5% less time on ChatGPT a month later than their average over the previous eight months. As of the second quarter to date, Claude had 56 million global monthly active users, while its year-over-year growth of about 640% far outpaced ChatGPT’s 62%, even though ChatGPT remains far larger overall.
Microsoft Releases Copilot Health Preview for US Microsoft 365 Subscribers With Wearable and Health Record Integration
Microsoft has moved Copilot Health into preview, giving eligible U.S. users aged 18 and older with Microsoft 365 Personal, Family, or Premium subscriptions access through Copilot on the web. The health-focused AI tool is designed to combine data from wearables, wellness apps, and medical records to provide personalized health insights, explain lab results, and help users prepare for doctor visits or find care providers. Microsoft said the service starts with support for Apple Health and can connect to records from more than 50,000 U.S. provider organizations, while responses are informed by trusted health sources and clinical input. The company also said Copilot Health chats are kept separate from the rest of Copilot, are not used to train AI, and the preview will expand gradually as feedback comes in.
Microsoft Launches Scout Personal Assistant for Microsoft 365, Built on OpenClaw Framework
Microsoft has launched Scout, a new AI assistant for Microsoft 365 that is inspired by OpenClaw and built on its framework. The cloud-based tool works across desktop and web, keeps a persistent identity, and can automate tasks using built-in and user-created skills tied to apps like email and calendars. Scout is being offered through Microsoft’s Frontier program and requires a GitHub Copilot subscription. The company said it has added security measures, including continuous policy checks and audit trails, to address concerns around autonomous AI agents, and the product was unveiled alongside other Build conference AI updates.
Microsoft Debuts ASSERT Framework for Text-Driven Testing of Application-Specific AI Behavior and Compliance
Microsoft has released ASSERT, an open-source framework designed to help developers test whether AI systems behave as intended in specific products and services. The tool converts plain-language descriptions of goals, policies, and expected behavior into structured test cases, then runs and scores those tests against the target system. It can also trace intermediate steps and tool calls, helping developers identify where failures occur and customize evaluations using system context, tools, and constraints. The release reflects a broader industry push toward repeatable AI testing and regression checks as companies look beyond general benchmarks to measure application-specific behavior.
Amazon Adds AI-Generated Product Images to Search Results, Raising Concerns Over Misleading Shopping Experience
Amazon said it will start showing AI-generated product images in its shopping app for some search queries to help people refine what they are looking for when they do not know the right terms. The images will appear below autocomplete suggestions and, when tapped, will lead shoppers to results that more closely match a style, such as different versions of a blue gingham dress. The move adds to Amazon’s wider push to use AI across shopping, including review summaries, visual search tools, shoppable collages, and voice-based shopping through Alexa. The feature may raise concerns that shoppers could mistake AI-made images for real products listed for sale, even though the visuals are meant only to guide searches.
Google Labs Launches Dreambeans AI App That Turns Personal Data Into Daily Cartoon Stories
Google Labs has released Dreambeans, a new AI app for Android and iOS that uses data from Google services such as Gmail, Calendar, Photos, YouTube, and Search History, with user permission, to create a small daily set of illustrated story cards. The app is designed to offer personalized lifestyle suggestions, including places to visit, topics to look into, events to note, and tips tied to upcoming plans, rather than endless social feeds. Google says users can choose which services to connect, delete their data, and that only the user can access the generated stories. Dreambeans is currently limited to eligible U.S.-based Google AI Ultra subscribers, while other users with personal Google accounts can join a waitlist.
Meta Rolls Out AI Creator Assistant on Facebook With Personalized Insights and Content Recommendations
Meta has rolled out an AI creator assistant on Facebook that gives creators personalized advice based on their content style, performance, audience, and goals. The tool can answer questions such as the best time to post, summarize comment feedback, track audience shifts over time, and suggest content ideas using trends like popular audio and cultural moments. It is currently being released to creators in the U.S., Canada, and India, as Meta looks to keep creators active on Facebook amid competition from TikTok and YouTube. The company also expanded AI translation features on Facebook with support for Arabic, Bahasa Indonesian, French, Thai, and Vietnamese, while saying more than 500 million users now watch AI-translated videos on the platform each week.
Apple Approves Poke as First Standalone AI Agent on Messages for Business Platform
Apple has approved Poke as the first standalone AI agent to operate on its Messages for Business platform, marking a shift for a service previously used mainly by brands such as airlines, retailers, and hotels to message their own customers. Poke, launched in March, lets users interact with an AI agent through text for tasks such as planning, calendar management, health tracking, smart home control, and photo editing, and the company said it has handled about 100 million messages so far. The service already works over SMS, Telegram, and in some markets WhatsApp, and it will now be available through Apple Messages for Business as a verified business account. The move comes just ahead of Apple’s WWDC event, where the company is widely expected to outline broader AI plans, though this approval does not mean Apple has opened its App Store or messaging platform broadly to consumer AI agents.
Anthropic Files Confidentially for IPO After $65 Billion Funding Round Lifts Valuation
Anthropic said it has confidentially filed with the U.S. Securities and Exchange Commission for a proposed initial public offering, a move that lets it prepare privately before releasing full financial details in a public S-1 filing. The company said the IPO’s timing will depend on market conditions and did not disclose share numbers or pricing. The filing follows a recent Series H fundraising round and comes as competition in the AI sector intensifies, with OpenAI also expected to pursue an IPO. Anthropic, founded in 2021 and known for its Claude models, has grown quickly in enterprise AI and recently said its revenue run-rate had climbed sharply, while it continues limited testing of its new Mythos model over cybersecurity concerns.
🎓AI Academia
Report Warns Misaligned AI Models Could Become Insider Threats in Government and Defense Networks
A new policy memo on arXiv argues that AI systems used in sensitive government and contractor settings should be treated as a new kind of insider risk, much like trusted human employees with privileged access. The paper says advanced models placed inside high-security environments could misuse authorized access to leak information, enable sabotage, theft, or other harmful actions if their behavior becomes misaligned with organizational goals. It warns that current insider-threat programs were built for people, not AI agents, even as these systems gain more autonomy and are increasingly deployed in high-stakes operations. The authors recommend adapting existing safeguards, including continuous monitoring and evaluation, to cover AI models working in classified and other critical environments.
Study Maps How Insurance Markets Are Adapting to Risks From Agentic AI Systems
A new paper argues that agentic AI is creating insurance risks that go beyond traditional categories because these systems can act on their own, use tools, make decisions, and even change digital or physical environments. It says the biggest threats include hallucinations, prompt-injection attacks, decision mistakes, model drift, third-party dependency failures, and cyber-physical damage. The analysis finds that insurers will likely need a mix of products, including cyber, tech errors and omissions, product liability, performance warranties, and dedicated AI liability coverage, rather than a single standalone policy. The paper also says better governance, transparency, telemetry, and clearer regulation will be critical as insurers work out how to price, underwrite, and manage large-scale losses tied to autonomous AI systems.
Huawei Researchers Detail Compositional Governance Framework for Authorization, Delegation, and Scope Control in Agentic AI
A new research paper argues that today’s access-control systems are not built for agentic AI, where software agents can act autonomously, delegate tasks to other agents, and operate across tools and organizations. The paper says common approaches such as OAuth 2.0 rely on fixed tokens and static scopes, which do not capture recursive delegation, time limits, contextual rules, or shifting authority in multi-agent systems. It proposes a compositional authorization framework that treats delegation as a runtime contractual condition and uses scope attenuation to ensure permissions narrow as they pass down a chain. The researchers say the model can overlay these agent-specific controls onto existing authorization systems without rewriting them, and they back the approach with formal proofs and empirical evaluation aimed at making agentic AI governance more accountable and practical.
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.
