Clarifai Deletes 3 Million OkCupid Photos Used to Train Facial Recognition AI
++ Platforms tighten AI media controls (Deezer, YouTube), major security and supply-chain alerts around Anthropic Mythos and vendor access, White House and Trump administration warn of AI theft..
Today’s highlights:
Clarifai has deleted 3 million photos that OkCupid allegedly shared in 2014 to help train facial recognition AI, according to Reuters, and said it also removed any models built using that data. The report cites an FTC investigation that found OkCupid provided user-uploaded images, along with demographic and location data, despite privacy policies that should have barred such sharing. The FTC began investigating in 2019 after reports that Clarifai had used OkCupid images to build tools estimating a person’s age, sex, and race from facial images. OkCupid and parent company Match Group settled with the FTC last month without admitting the allegations, but are now barred from misrepresenting or helping others misrepresent how user data is collected and shared.
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
White House Memo Alleges Chinese Firms Conduct Large-Scale Theft of US Artificial Intelligence Technology
The White House says it will step up coordination with US AI companies after an internal memo claimed foreign groups, mainly in China, are carrying out large-scale efforts to copy American AI technology through a technique known as distillation. The memo says these actors use thousands of accounts to probe AI systems, extract useful information, and apply it to their own model development, which the administration describes as an attempt to weaken US research and gain proprietary data. Officials said the government will share more threat intelligence, improve coordination with companies, develop best practices, and consider ways to hold foreign actors accountable, though no specific penalties were announced. China’s embassy rejected the claims, saying Chinese innovation comes from domestic effort and international cooperation, while companies including OpenAI and Anthropic have previously alleged similar activity involving Chinese AI labs such as DeepSeek, Moonshot and MiniMax.
GRAI Raises $9 Million to Build Social AI Music Tools With Artist Controls
GRAI, a new AI music startup backed by a $9 million seed round, is betting that consumers want to interact with music by remixing, sharing, and modifying songs rather than generating tracks from scratch. The company says its focus is on making music more social while giving artists and labels control over whether their songs can be used, with an opt-in or opt-out model. GRAI has already released early consumer apps on iOS and Android to test how people, especially Gen Z and Gen Alpha users, want to engage with music beyond passive listening. The startup is also building audio systems designed to preserve the identity of original tracks while enabling legal transformations that could create new royalty opportunities for rights holders.
Report Says Unauthorized Group Accessed Anthropic’s Exclusive Mythos Cybersecurity Tool Through Third-Party Vendor
Anthropic is investigating reports that an unauthorized group gained access to its restricted cybersecurity model, Mythos, through a third-party vendor environment, according to Bloomberg and a company statement to TechCrunch. The company said it has found no evidence so far that its own systems were affected, but the group reportedly used the tool regularly and showed screenshots and a live demo as proof. Bloomberg said the users are linked to a Discord community focused on unreleased AI models and may have located Mythos by guessing its online address format on the day it was unveiled. Mythos was shared only with a small set of partners under Project Glasswing because Anthropic has warned the tool could be misused for hacking if it falls into the wrong hands.
Meta to Record Employee Keystrokes and Mouse Movements to Train AI Models Internally
Meta plans to collect some employees’ keystrokes, mouse movements, clicks, and navigation patterns on certain internal applications to help train AI systems, according to a Reuters report and a company statement to TechCrunch. The company said the goal is to give its models real examples of how people complete everyday computer tasks, while adding that safeguards are in place to protect sensitive content and that the data will not be used for other purposes. The move highlights how AI companies are searching for new sources of training data as they work to build more capable and efficient models. It also raises fresh privacy concerns, coming amid broader reports that companies are increasingly turning internal workplace data into material for AI training.
Delve Faces Fresh Scrutiny After Context AI Breach and Lovable Security Failures Emerge
TechCrunch confirmed that troubled compliance startup Delve handled security certification for Context AI, whose recent security incident was linked to the breach at Vercel. Context AI said it has since dropped Delve, moved its compliance work to Vanta, and hired an independent auditor for new examinations. Separately, Lovable, another former Delve customer, said it had already stopped using Delve but still disclosed its own incident involving public exposure of customer chat data due to a configuration error. The developments add to Delve’s growing scrutiny after whistleblower allegations over its certification practices, the loss of customers, and fresh unverified claims about its business conduct.
US Justice Department Backs xAI Lawsuit Challenging Colorado Artificial Intelligence Regulation Law
The U.S. Justice Department has joined xAI’s lawsuit against a Colorado law that would regulate certain “high-risk” AI systems used in areas such as jobs, housing, healthcare, education, and finance. The department argued the law may violate the Constitution’s equal protection guarantee by requiring companies to prevent discriminatory harms while allowing some diversity-related distinctions. xAI separately argues the law violates the First Amendment by restricting how AI systems are designed and by compelling speech. The Colorado law is set to take effect on June 30, and the federal intervention raises the dispute into a broader clash over whether AI should be regulated state by state or through a single national framework.
Trump Administration Vows Crackdown on Chinese Firms Accused of Exploiting US AI Models
The Trump administration said it will step up action against foreign companies, especially those in China, accused of extracting capabilities from U.S.-made AI models through techniques such as distillation. A White House memo said the government will work with American AI firms to detect misuse, strengthen defenses, and consider penalties for offenders, as Washington argues the U.S. must stay ahead in AI for economic and military reasons. The move comes as reports show China has rapidly narrowed the performance gap with the U.S., while Chinese officials reject the accusations and call them an attempt to suppress China’s tech sector. The push also aligns with bipartisan support in the House for a bill that would create a process to identify and punish foreign actors accused of copying key features of closed-source American AI systems.
FM Nirmala Sitharaman Meets Bank Chiefs to Review AI Risks After Anthropic Mythos Concerns
Finance Minister Nirmala Sitharaman on Thursday met heads of banks, RBI officials, and the Ministry of Electronics and Information Technology to review risks that artificial intelligence could pose to India’s financial system. The discussion followed global concerns around Anthropic’s Claude Mythos model, which the company says can identify and exploit serious software vulnerabilities and has been withheld from public release because of cybersecurity risks. Officials said banks were asked to take preemptive steps to protect systems, customer data, and funds, while the finance ministry and RBI assess the extent of any possible threat. A senior official said Indian financial systems remain secure for now and there is no immediate cause for undue concern, even as regulators continue due diligence.
Meity Proposes Continuous AI Watermarking After Disappointment Over Compliance With Existing Rules
India’s Ministry of Electronics and Information Technology is considering a move to require “continuous” watermarking of AI-generated content, amid dissatisfaction with how platforms have followed earlier advisory rules on labeling synthetic media. The proposal is aimed at making AI-generated text, images, audio and video easier to identify throughout their lifecycle, rather than through one-time or easily removable labels. The move reflects the government’s broader push for stronger accountability from AI firms as generative tools spread rapidly and concerns over misinformation and deepfakes grow. If adopted, the measure could add stricter compliance demands on AI platforms operating in India and shape future rules on the traceability of synthetic content.
Deezer Reports 44% of Daily Song Uploads Are AI-Generated as Platform Tightens Controls
Deezer said AI-generated songs now make up 44% of all new tracks uploaded to its platform, with nearly 75,000 such songs arriving each day, or more than two million a month. While AI music uploads are rising sharply from about 10,000 daily in January 2025, actual listening remains limited at 1% to 3% of total streams, and the company said 85% of those streams are flagged as fraudulent and stripped of monetization. Deezer has responded by removing AI-tagged tracks from algorithmic recommendations and editorial playlists, and it said it will stop storing hi-res versions of those songs. The update comes as AI music gains wider visibility across the industry, including chart success on iTunes, while surveys cited by Deezer suggest most listeners cannot reliably distinguish AI-made songs from human-created music and want clearer labeling.
YouTube Expands AI Likeness Detection Tool to Celebrities, Talent Agencies, and the Entertainment Industry
YouTube is widening access to its AI “likeness detection” tool, extending it to celebrities and the entertainment industry after earlier pilots with creators and a broader rollout to politicians, government officials, and journalists. The system works like Content ID, but for simulated faces, helping detect AI-generated videos that use a person’s likeness without permission. Talent agencies and management firms including CAA, UTA, WME, and Untitled Management have backed the effort, and enrolled participants do not need their own YouTube channels to use it. When a match is found, users can request removal, file a copyright claim, or take no action, though parody and satire may still be allowed under YouTube’s rules. YouTube also said the tool will later add audio detection and continues to support the NO FAKES Act, while removals tied to the system remain very limited so far.
🚀 AI Breakthroughs
OpenAI releases GPT-5.5 as latest model, advancing push toward an AI super app
OpenAI on Thursday released GPT-5.5, describing it as its smartest and most intuitive model so far and positioning it as another step toward a unified AI “super app” that could combine ChatGPT, coding tools, and browsing features. The company said the model improves performance in enterprise work such as agentic coding and knowledge tasks, while also showing gains in mathematics, scientific research, and technical workflows, including potential use in drug discovery. OpenAI also claimed GPT-5.5 outperforms earlier in-house models and rival systems from Google and Anthropic on a range of benchmarks, though those results were presented by the company itself. GPT-5.5 began rolling out Thursday to Plus, Pro, Business, and Enterprise users in ChatGPT, with GPT-5.5 Pro available for Pro, Business, and Enterprise customers.
ChatGPT Images 2.0 Shows Major Gains in Text Rendering and Complex Image Generation
OpenAI’s new ChatGPT Images 2.0 model marks a notable improvement in AI image generation, especially in rendering readable text, a task that older diffusion-based systems often failed at. The company said the model has “thinking capabilities” that help it search the web, check its work, and create more complex outputs such as marketing assets in multiple sizes and multi-panel comics. OpenAI also said the system is better at handling non-Latin scripts including Japanese, Korean, Hindi, and Bengali, though its knowledge cutoff is December 2025. Images 2.0 is rolling out to all ChatGPT and Codex users, with paid tiers getting higher-end generation options, while the gpt-image-2 API will be offered with pricing based on image quality and resolution.
OpenAI Launches Codex-Powered Workspace Agents in ChatGPT for Teams and Enterprise Workflows
OpenAI has launched workspace agents in ChatGPT, a new Codex-powered feature that lets teams build shared AI agents to handle multi-step work such as reporting, coding, lead follow-up, software reviews, and vendor risk checks. The agents run in the cloud, can work across connected tools including ChatGPT and Slack, support memory and scheduled tasks, and can be shared across teams with organization-level permissions and approvals. The company said the feature is available in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans, while existing GPTs will remain available and will later be convertible into workspace agents. OpenAI also said the product includes admin controls, analytics, Compliance API visibility, and prompt-injection safeguards, and that it will remain free until May 6, 2026 before shifting to credit-based pricing.
DeepSeek Previews V4 AI Models, Claims Narrower Gap With Frontier Systems at Lower Cost
Chinese AI lab DeepSeek has previewed two new open-weight large language models, DeepSeek V4 Flash and V4 Pro, saying they narrow the performance gap with today’s frontier AI systems. The company said both are mixture-of-experts text-only models with 1 million-token context windows, while V4 Pro has 1.6 trillion total parameters, making it larger than rival open-weight models from Moonshot AI and MiniMax. DeepSeek claimed the models improve on V3.2 in efficiency, reasoning, and coding, with benchmark results that in some cases approach or surpass leading systems from OpenAI and Google, though they still trail top models on some knowledge tests by an estimated three to six months. The company also positioned V4 as significantly cheaper than many frontier competitors, even as its release comes amid wider U.S. accusations that Chinese firms have used proxy accounts and model distillation to copy American AI technology.
Microsoft Copilot Agentic Features in Word, Excel and PowerPoint Become Generally Available Today
Microsoft said Copilot’s new agentic features in Word, Excel, and PowerPoint are now generally available, allowing the AI assistant to take multi-step actions directly inside documents, spreadsheets, and presentations instead of only answering prompts. The company said the upgrade is powered by stronger AI models and Work IQ context, helping Copilot handle tasks such as rewriting documents, editing formulas and tables, and updating slide decks while keeping users in control of changes. Microsoft also shared early usage data showing higher engagement, retention, and satisfaction across the three apps, with the biggest gains reported in Excel. The features are now the default experience for Microsoft 365 Copilot and Microsoft 365 Premium users, and are also available to Microsoft 365 Personal and Family subscribers.
Kimi K2.6 Open-Sourced With Stronger Coding, Agent Swarms, and Long-Horizon Task Performance
Moonshot AI has open-sourced Kimi K2.6, a new AI model focused on coding, long-running agent tasks, and multi-agent “swarm” workflows, and made it available through its app, website, API, and coding tools. The company says the model improves sharply over K2.5 in long-horizon software engineering, citing internal tests and enterprise beta feedback that point to stronger tool use, better instruction following, and more reliable performance across large codebases, front-end work, and DevOps tasks. In benchmark tables published in the technical blog, K2.6 posts competitive results against leading closed models on agentic and coding tasks such as SWE-Bench Pro, Terminal-Bench 2.0, and HLE with tools, while also showing gains in search, reasoning, and some vision evaluations. The post also highlights newer “agent swarm” features that can split jobs across up to 300 sub-agents and a research preview called Claw Groups, aimed at persistent, collaborative AI agents that can work across devices and applications over extended periods.
Bond Launches AI-Powered Social Platform Aimed at Reducing Doomscrolling Through Real-World Recommendations
Bond, a new social media app that launched Tuesday, says it wants to reduce doomscrolling by using AI to turn users’ posts, photos, videos, and audio memories into personalized real-world recommendations, such as restaurants, events, and activities. Unlike traditional platforms built around endless feeds, Bond has no main feed and instead centers on profile-based story updates that disappear publicly after 24 hours but remain in a private archive. The company says its long-term business model could include letting users license their stored memories for AI training or using the data for opt-in commerce recommendations, rather than selling ads. Bond also says users can delete memories and profiles, but its end-to-end encryption is not yet in place and remains a future priority.
Google Expands AI Overviews to Workplace Gmail and Drive for Workspace Customers
Google said at its Cloud Next conference that AI Overviews are coming to Gmail for workplace users, allowing them to ask natural-language questions in search and get concise answers drawn from multiple emails and conversations without opening each message. The feature is designed to help with common work queries such as project milestones, invoices, trip details, deck feedback, and performance updates. It will be enabled by default when Gemini for Workspace in Gmail and Workspace Intelligence access to Gmail are turned on, with additional smart feature settings also required for end users. Google also said the capability, previously limited to Google AI Pro and Ultra consumer plans, is expanding to eligible business, enterprise, education, and Frontline customers, while AI Overviews in Drive are now becoming broadly available after beta.
Google Targets IT Teams With Gemini Enterprise Agent Platform for Large-Scale Enterprise Agent Management
Google used its Cloud Next event to detail Gemini Enterprise Agent Platform, a new tool for enterprises to build and manage AI agents at scale, positioning it against Amazon Bedrock AgentCore and Microsoft Foundry. The company is aiming the platform mainly at IT and technical teams, reflecting the stronger traction of AI agents in coding and other technical work, as well as ongoing enterprise concerns around security. Business users are being steered to the Gemini Enterprise app, where they can use IT-built agents or create their own for tasks such as scheduling, automating routine workflows, and editing files across apps. Google also said these tools can run on multiple models, including its own Gemini and Nano Banana 2, alongside Anthropic’s Claude family, with support for Claude Opus, Sonnet, Haiku, and the newly released Opus 4.7.
UAE Targets Shifting 50% of Government Services to AI Within Two Years
The United Arab Emirates plans to move 50% of government services, sectors and operations to AI within the next two years, expanding its push to modernize the public sector. The strategy will use autonomous agentic AI systems to handle routine work, analyze data and support decision-making with limited human involvement. Federal agencies will be measured on how quickly they adopt AI, redesign workflows and integrate smart systems, while government employees will receive AI-focused training. Officials say the goal is to reduce bureaucracy, lower costs and improve service speed, building on earlier digital reforms such as UAE Pass and Government Services 2.0, as well as Abu Dhabi’s separate target to become a fully AI-native government by 2027.
NSA Reportedly Uses Anthropic’s Mythos Despite Pentagon Dispute Over Access and Supply-Chain Risk
The National Security Agency is reportedly using Mythos Preview, Anthropic’s restricted cybersecurity model, even as the Pentagon has labeled the company a supply-chain risk amid a dispute over access to its AI systems. Axios said Mythos was not released publicly because Anthropic believed it could enable offensive cyberattacks, and access was limited to about 40 organizations. The NSA is said to be using the model mainly to scan systems for exploitable vulnerabilities, while the U.K.’s AI Security Institute has also confirmed access. The development highlights a tension in Washington, where the U.S. military is expanding its use of Anthropic’s tools while also arguing that those same tools could pose national security risks.
🎓AI Academia
OpenAI Releases Privacy Filter, a 1.5B-Parameter Open-Weight Model for Masking Personal Data in Text
OpenAI has released Privacy Filter, an open-weight 1.5B-parameter model designed to detect and redact personally identifiable information in text, including names, addresses, emails, phone numbers, dates, account numbers, and secrets such as passwords or API keys. The company said the model is built for high-throughput privacy workflows, can run locally for on-device redaction, supports up to 128,000 tokens, and is available under the Apache 2.0 license on Hugging Face and GitHub for commercial use and fine-tuning. OpenAI reported that Privacy Filter scored 96% F1 on the PII-Masking-300k benchmark, rising to 97.43% on a corrected version of the dataset after annotation issues were reviewed. The company said the model is meant to strengthen privacy protections in AI pipelines such as training, logging, indexing, and review, but noted that it is not a full anonymization or compliance solution and may still require human review in sensitive domains.
Study Proposes Statistical Certification Framework to Quantify and Verify Acceptable Risk in AI Regulation
A new paper argues that AI regulation still lacks a practical way to measure whether high-risk systems are actually safe enough before deployment, even as rules like the EU AI Act and the NIST AI Risk Management Framework demand such proof. It proposes a two-step certification model: regulators would first set a clear acceptable failure rate and define the operating conditions, then statistical tools called RoMA and gRoMA would calculate an auditable upper bound on the system’s real failure risk without needing access to the model’s internal workings. The paper says this could help turn broad legal requirements into a concrete compliance process for opaque AI systems such as neural networks. It also argues that the approach could shift accountability more clearly onto developers by producing measurable safety evidence that fits within existing legal frameworks.
Cornell Survey Examines Agentic Artificial Intelligence Applications, Risks, and Regulation Across Finance
A new survey posted on arXiv examines how agentic AI could reshape finance by focusing on systems that can reason, plan, and make adaptive decisions with limited human input. The paper reviews how these autonomous AI tools may be used across trading, portfolio management, risk analysis, compliance, and other financial operations, while also looking at the system designs that support them. It also highlights regulatory and governance questions, including oversight, accountability, and the broader market risks that could come from increasingly autonomous decision-making. Overall, the survey frames agentic AI as a major shift for financial markets, with potential efficiency gains alongside serious concerns about control, transparency, and systemic impact.
Study Finds Internal Expert Collaboration Helps Companies Turn EU AI Act Rules Into Development Practice
A study to be presented at ACM FAccT 2026 examines why turning the EU AI Act into day-to-day software practice remains difficult, especially inside smaller AI companies. Based on insider action research at an AI startup, the paper outlines a legal-text-to-action process that turns regulatory requirements into concrete tasks through internal collaboration between legal, product, and technical experts. It finds that teams respond to rules in three main ways: some requirements match existing development goals, some are already covered by current practice, and others are dismissed as administrative overhead. The paper argues that governance efforts are taken more seriously when developers see a clear benefit for product quality or user protection, while verification-heavy requirements are more likely to become box-ticking exercises.
Study Proposes Simple AI Incident Trajectory Classification to Separate Reporting Bias, Exposure, and Harm Trends
A new preprint argues that headline counts of AI incidents can be misleading because they mix together media reporting trends, wider AI deployment, and the actual rate of harm. The paper says public incident databases, which are largely built from news reports, tend to overrepresent dramatic, English-language cases while missing slower, systemic harms. To address this, it proposes a simple classification framework that separates exposure from harm rates, uses proxy measures when data is limited, and avoids false precision by focusing on directional trends. The authors say the approach can help policymakers judge whether AI risks are truly rising or simply being reported more often, while also showing where better reporting systems are still needed.
Study Finds Structural Quality Gaps in AI Governance Prompts Across Public AGENTS.md Files
A new study examines 34 publicly available AGENTS.md files on GitHub and finds that many practitioner-written AI governance prompts are structurally incomplete. Using a five-part evaluation framework drawn from requirements engineering and related theory, the paper reports that 37% of file-model pairs fell below its completeness threshold. The most common missing elements were data classification rules and clear assessment rubrics, two basics needed to define what an AI agent can handle and how its output should be judged. The study argues that weak prompt structure, rather than model capability alone, can leave AI systems misaligned with organizational intent, and says these gaps could be detected and fixed through automated static analysis tools.
MedSkillAudit Study Proposes Domain-Specific Framework to Audit Medical Research Agent Skills Before Deployment
Researchers from AIPOCH and Zhongshan Hospital at Fudan University have proposed MedSkillAudit, a domain-specific framework designed to check whether medical research AI agent skills are safe and reliable enough for release. In a test covering 75 skills across five medical research categories, the system measured quality, release readiness, and high-risk failures, with results compared against reviews from two human experts. The study found that 57.3% of the skills did not meet the threshold for limited release, highlighting broad quality and safety gaps before deployment. MedSkillAudit showed moderate agreement with expert consensus and slightly outperformed human inter-rater consistency on scoring, though it struggled in academic writing tasks, where the rubric did not align well with expert judgment.
Microsoft Research Preprint Finds LLMs Corrupt Documents Across Long Delegated Editing Workflows
A Microsoft Research preprint under review says current large language models can silently damage documents when used for delegated editing work over many steps. In a new benchmark called DELEGATE-52, which tests long workflows across 52 professional domains such as coding, crystallography, music notation, and 3D files, 19 models were evaluated on repeated document edits. The paper reports that even top models including Gemini 3.1 Pro, Claude 4.6 Opus, and GPT-5.4 corrupted about 25% of document content on average after 20 interactions, while some other models performed worse. It also says tool-using agents did not improve results, and that larger documents, longer sessions, and extra distractor files made the failures more severe, raising concerns about trusting LLMs with complex editing tasks.
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.
