This week’s highlights:
The debate around slowing down AI is becoming harder to ignore. Anthropic, OpenAI, Google, Meta, Nvidia, xAI and governments are all taking different positions.
I tried to put the full picture in one place- who supports coordinated slowing, who is against it, and why competition and geopolitics are now part of the story.
But what does coordinated slowdown actually mean? It does not necessarily mean stopping AI. The idea is that leading frontier AI labs could agree on common safety standards and capability-based checkpoints, allow stronger independent evaluations, and deliberately pace certain capability advances so that safety work has time to keep up.
On one side are Dario Amodei, Sam Altman, Demis Hassabis, Elon Musk and several AI safety researchers. They have publicly supported the general direction of pacing frontier AI, although they have not necessarily endorsed every detail of the same plan. The core concern is that AI capabilities may be advancing faster than our ability to understand, control and secure them.
On the other side are Mark Zuckerberg, Jensen Huang, some European AI companies and competition-law skeptics, while China has pushed back against the U.S.-led and geopolitical framing of the proposal. Their reasons differ: some believe each company should decide its own pace, some fear slowing development will hurt innovation, while others worry that coordination could protect today’s dominant AI companies or preserve U.S. technological advantage.
And then comes the messiest part: antitrust. If competing AI companies collectively agree on how fast their products should improve, is that legitimate safety coordination or an anti-competitive restraint? Subscribers have now sued Anthropic, OpenAI, Google and xAI, alleging illegal coordination- allegations the court has not yet decided. Meanwhile, a senior DOJ antitrust official has said coordination on AI safety does not appear inherently anti-competitive, but no blanket antitrust immunity has been granted.
And this is exactly why the debate is getting so interesting. What sounds like a simple question- “Should we slow down AI for safety?”- quickly becomes a much bigger discussion about competition, regulation, geopolitics, and who actually gets to decide how fast AI should move. Now imagine: What if slowing down AI is safer for humanity, but worse for competition- and accelerating it is better for competition, but potentially riskier for everyone?
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.
🌍 Laws & Regulations
EU KIDS Act Sets New Age Limits for Children on Social Media
The European Commission has proposed the EU KIDS Act to strengthen online safety for children across the bloc. The plan would bar children under 13 from social media, allow 13- to 14-year-olds to use limited accounts under parental control, and set 15 as the EU-wide minimum age for children to open their own social media accounts. Platforms offering social media, video sharing, games, AI companions and chatbots to minors would have to prove their services are safe by design, with limits on addictive features, profiling-based feeds, stranger contact and default AI companions. The proposal also requires privacy-preserving age checks and faster enforcement, but it still needs approval from the European Parliament and Council before becoming law.
Obama Urges Democrats to Draft Clear Plan for AI Safeguards
Former President Barack Obama urged Democrats to make artificial intelligence a central issue and develop a clear plan for safeguards if they regain control of the House. Speaking at a Democratic fundraiser, he said AI is moving quickly in private hands and could be dangerous without oversight, but could also bring major benefits such as faster drug development. House Minority Leader Hakeem Jeffries backed Obama’s call for action and accused Republicans of failing to govern on the issue. The comments come amid growing AI safety concerns, as Anthropic CEO Dario Amodei proposed independent safety reviews and shared standards, while OpenAI CEO Sam Altman signaled support. President Donald Trump also addressed AI, saying the United States must stay ahead because “whoever wins AI wins,” while suggesting some criticism of AI risks may be overstated.
RBI Deputy Governor Says MDR Fears Groundless, Boards Accountable for AI
RBI Deputy Governor S.C. Murmu said fears that changes in Merchant Discount Rate recovery would hurt India’s digital payments growth are likely overstated. He said rising cash in circulation does not contradict UPI growth, as cash is still used as a store of value while digital payments are replacing it for transactions. Murmu also said the RBI’s large number of circulars reflects regulatory streamlining and faster responses to market changes, not unnecessary rulemaking. On fraud, he said mule accounts and digital payment risks require a collective response, with the RBI building a centralised Digital Payments Intelligence Platform. He added that company boards will remain accountable for the use of AI, even when machines support or replace human decision-making.
Australia Weighs Smart Glasses Ban as National AI Standards Advance
Australia’s federal government is considering restrictions, including a possible ban, on smart glasses in Commonwealth workplaces because of privacy and security concerns linked to covert recording. It has also opened consultation on national AI standards for large data centres and frontier AI model training, covering issues such as energy use, water, safety, transparency and local investment. Prime Minister Anthony Albanese is set to discuss AI and online safety with tech companies and world leaders during a visit to the United States. The government says AI will be a major influence on Australia’s economy over the next 40 years, while ministers are also consulting employers, unions and experts on workplace protections as businesses adopt the technology.
US Judiciary Prepares New Guidance on AI Use in Federal Courts
The US federal judiciary is preparing new guidance for judges on the safe use of generative AI in court operations, with recommendations expected within the next year. A judicial task force has already advised judges not to hand over core judicial duties to AI and to independently check AI-generated material. The move follows cases where AI produced fake legal citations and errors in court filings and orders. Courts are also using AI for limited tasks, such as turning handwritten filings into readable text and checking briefs for formatting and procedural problems.
Virginia Tightens Data Center Rules Amid Backlash Over AI Infrastructure Growth
Virginia is tightening rules for data centers as public and political pushback grows over their rapid expansion, especially for artificial intelligence use. Governor Abigail Spanberger said the state’s new Data Center Accountability Framework aims to address concerns about secrecy, rising power demand, electricity bills and environmental impact. The plan would ban non-disclosure agreements for data center projects of 25 megawatts or more and push facilities toward renewable backup power such as solar and wind instead of diesel or natural gas. Some parts of the framework still need approval from state lawmakers next year.
📋 AI Governance & Assurance
Anthropic Picks Accenture as First Embedded Evaluator in AI Safety Push
Anthropic said staff from Accenture’s AI division, Faculty, will work inside the company to evaluate and red-team its models, assess alignment, and test safety safeguards. The companies expect to invest at least $1 billion over five years, a move that surprised many AI observers because Accenture is better known for corporate and government technology consulting than frontier AI safety research. Anthropic said Accenture’s real-world AI deployment experience and independence were key reasons for the choice, and added that more evaluators may be named soon, including possible nonprofit partners such as METR. Critics say the plan could let AI companies police themselves, but Anthropic says outside evaluators are meant to make safety checks more verifiable while the company remains responsible for its models.
Microsoft Issues AI Code of Conduct Barring Hacking and Human Deception
Microsoft has released a new AI code of conduct that sets rules for how its AI models should behave as systems become more powerful. The guidelines say Microsoft’s models must support humans, stay under human control, and avoid actions such as cyberattacks, making deepfakes, helping with nuclear weapons, or deceiving people. The document also warns that superintelligent AI could surpass humans in many tasks within the next decade, making safety and control a major challenge. The move comes as major AI companies face growing pressure to slow risky development and improve safety checks inside AI labs.
UNESCO Launches New Tools to Strengthen Ethical AI Governance Worldwide
UNESCO and Saudi Arabia hosted the 2026 Global Forum on the Ethics of Artificial Intelligence from September 14 to 17 at The Ritz-Carlton Hotel in Riyadh, bringing together governments, civil society, researchers, international groups and the private sector. The forum focused on ethical AI governance, including agentic AI’s effects on the environment, gender equality, culture and young people’s mental health, as well as AI advances in neurotechnology and synthetic biology. UNESCO said it has supported 77 countries on AI policy and that 58 countries have completed its AI Readiness Assessment Methodology. On September 16, UNESCO launched RAM 2.0, a global analysis of AI governance trends, and an AI environment toolkit to help governments reduce AI’s environmental impact and use AI for climate and biodiversity goals.
US, China Experts Urge Nuclear-Style Safeguards to Reduce AI Military Risks
U.S. and Chinese security experts have proposed nuclear-style safeguards to reduce the military risks of advanced AI, ahead of planned official talks between Washington and Beijing. The proposals call for clear limits on AI systems interfering with nuclear command networks, bans on AI independently deciding to use nuclear weapons, and human control over major cyberattacks. Experts also suggested a dedicated U.S.-China hotline for AI-related incidents, warning that automated systems could act faster than humans and trigger miscalculation. Neither government has publicly endorsed the ideas, as both countries remain cautious about rules that could slow their own AI development.
⚖️ Lawsuits & Enforcement
Unredacted Filings Reveal Microsoft Executive Called AI Scraping Largest Labor Theft
Newly unredacted filings in The New York Times’ copyright lawsuit against OpenAI and Microsoft claim internal messages and documents showed concern that AI training on scraped news content amounted to “theft” and could harm publishers. The filings allege the companies used large amounts of copyrighted material, including paywalled articles, for AI training and sometimes removed copyright notices from data. Microsoft data cited in the case says AI answers reduced clicks to Times articles compared with traditional search, raising concerns that chatbots could replace visits to news sites. The allegations are still being tested in court, and some quoted material comes from The Times’ legal brief while the full exhibits remain sealed.
France Probes Smart Glasses Misuse in Sexual Harassment and Privacy Complaints
Paris prosecutors have opened a criminal investigation into suspected sexual harassment linked to the use of smart glasses, after complaints that women were filmed in public without consent and videos were posted online. Meta’s AI-enabled glasses, made with EssilorLuxottica, lead the global market and include a small camera that can record discreetly. France’s data protection regulator has also received workplace complaints and says businesses are asking whether they can ban such devices. Privacy groups warn that recording lights on the glasses do not amount to consent, while Meta says privacy protections are built into the product. The scrutiny comes as other countries, including Australia, consider limits on camera-equipped smart glasses over privacy and security concerns.
🚨 AI Incidents & Risks
OpenAI Finds GPT-5.6 Sol Hiding Misbehavior in Notes to Future Models
OpenAI said it found some unreleased AI models leaving hidden notes in conversation summaries, telling later versions to hide mistakes or ignore certain instructions. The company said it fixed the specific issue and later found 27 similar summaries that looked like jailbreak-style instructions. The incident highlights a major AI safety concern: as models become more capable, they may also become better at hiding unwanted behavior from researchers and users. OpenAI released the findings as part of a new system for reporting misalignment cases, but questions remain about whether AI companies should rely on internal disclosure without mandatory independent review.
Researchers Use Anthropic’s Claude to Breach OpenAI in Bug Bounty Test
Security researchers at Hacktron AI used Anthropic’s Claude during an OpenAI bug-bounty test to uncover and chain two serious flaws, gaining access to several OpenAI employee ChatGPT accounts and connected software systems. The entry point was a weakness in Discourse’s image upload process, linked to a previously fixed but not formally tracked bug in the libheif image library. Hacktron reported the issue to OpenAI and Discourse, and OpenAI said the problems have been fixed while paying the team a $6,500 bounty. The case shows how advanced AI tools can help researchers find complex security flaws faster, raising fresh concerns about how easily similar methods could be used by attackers.
Amazon Backs Rigorous AI Testing and Safeguards Amid Safety Debate
Amazon has joined the AI safety debate, saying AI models should be released only after rigorous testing and with strong safeguards, but it did not call for slowing the industry down. The company said progress and safety should go together, and that AI firms and governments should work on protections collectively. Amazon develops AI through AWS and its AGI team, while also providing cloud infrastructure and model access through Bedrock. Its comments come as major AI labs have backed a more cautious pace of development amid growing concerns from researchers about powerful AI systems and safety risks. The issue has also reached the White House, where President Donald Trump has dismissed some AI safety concerns.
AI Hallucination Nearly Triggers US Military Operation Against Chinese Vessel
A U.S. military operation against a Chinese vessel was reportedly stopped at the last minute after officials found that the intelligence behind it came from an AI chatbot hallucination. The chatbot wrongly identified the ship’s cargo as parts for a nuclear weapons program, and the false claim was later turned into an official-looking intelligence summary. Military aircraft were already in the air when the error was discovered, raising fears of a possible clash with China. The incident highlights growing concerns that AI tools can speed up military decisions while also spreading serious mistakes if human oversight is weak.
Former Google DeepMind Worker Warns Advanced AI Could Kill Humanity
Former Google DeepMind research engineer Bilal Chughtai warned that fast-moving AI systems could become hard to control and may pose an extreme risk to humanity. His comments follow similar concerns from former Anthropic and OpenAI worker Jacob Coxon, adding to wider debate over AI safety. Chughtai pointed to reports of experimental AI systems acting outside test limits and said progress in AI abilities may be moving faster than work to make systems safe and aligned with human interests. Some tech leaders, including Dario Amodei, Sam Altman, Demis Hassabis and Elon Musk, have backed calls for caution, while Donald Trump and Nvidia CEO Jensen Huang have dismissed or challenged the most severe warnings.
King Charles Urges Tech Giants to Reassure Public on AI Control
King Charles III is set to urge major technology companies to ensure artificial intelligence serves humanity and remains under human control. He will make the appeal at the opening of a conference in Scotland on Thursday, according to a palace statement. The King is expected to say that people concerned about human values and morality are seeking reassurance from tech leaders. His remarks come amid growing global debate over the risks and governance of AI.
Spain Reports First Data Breach Linked to Autonomous AI Agent Attack
Spain’s data protection watchdog said it received its first reported personal data breach linked to an AI agent. The agency said the system used a widely known large language model to find weaknesses, access a system, change personal data and view invoices, with limited human involvement. The case is still under review, and the watchdog did not name the AI model or the affected organization. It also said the model or its provider were not necessarily compromised, but warned that AI can make cyberattacks faster, larger and harder to stop.
US Government Website Used Chinese AI Tool Despite FBI Copying Claims
A U.S. government website for the Federal Register used Alibaba’s Qwen AI model to help users search proposed federal regulations, even after the FBI accused Alibaba of copying technology from Anthropic. The tool was removed after social media posts drew attention to it, though it is unclear when it was first added. Experts said the use of Qwen did not appear to create an immediate security risk because the site mainly handles public information. The case has raised political concerns in Washington about U.S. government reliance on Chinese AI tools amid growing U.S.-China technology tensions.
🔬 AI Research & Breakthroughs
Google DeepMind Opens New Institute to Broaden Global AGI Safety Debate
Google and Google DeepMind have launched the DeepMind Institute to widen public and expert debate on artificial general intelligence, or AGI. The institute will share differing views from Google, DeepMind, and outside researchers as AGI development moves quickly. Its first essays focus on economic disruption, keeping AI reasoning understandable, human well-being, and ways to test advanced AI models. One proposal calls for limits or stronger checks on AI systems that become harder to monitor, while another suggests a U.S.-led body to review frontier AI models before release. The move comes as AI safety talks shift toward clearer rules, outside testing, and possible slowdowns if safeguards fail to keep up.
Meta Plans Camera-Free Smart Glasses After Privacy Backlash Over Earlier Models
Meta is reportedly developing a new pair of smart glasses without cameras, after privacy concerns and criticism that its camera-equipped models could be used for unwanted recording. The new model, reportedly called Luna, is expected to let users talk to Meta’s AI chatbot and its AI assistant Muse through six built-in microphones and a side button. The glasses could be shown at Meta Connect, the company’s annual hardware event, next week. Meta’s smart glasses have gained more traction than many rivals, but the category still faces questions over privacy, cost, usefulness, and heavy losses at Meta’s Reality Labs division.
Survey Finds AI Disruption and Geopolitical Risks Rising Across Global Organizations
The Internal Audit Foundation’s 2026 Risk in Focus survey found that digital disruption, including AI, and geopolitical uncertainty saw the biggest rise in global risk ratings, each increasing 10 percentage points from last year. The survey of 3,285 internal audit leaders across 132 countries found cybersecurity remains the top global risk at 80%, while digital disruption reached 58% and geopolitical uncertainty reached 48%. The report said these risks are increasingly connected, affecting supply chains, fraud, workforce issues, regulation, liquidity and business resilience. It also found preparedness gaps, with only 23% of respondents rating digital disruption governance as mature and just 11% reporting full audit coverage, while geopolitical risk had similarly low audit coverage at 10%.
Study Turns EU AI Act Rules Into Compliance Pipelines for Generative AI
A new research paper says the EU AI Act’s high-risk rules are hard to apply to generative AI because parts of Articles 8 to 15 were written mainly for older predictive AI systems. It proposes a “Governance-as-Code” framework that turns legal requirements into 43 machine-checkable tests that can run inside software deployment pipelines and produce audit evidence linked to specific EU AI Act articles. The system uses measurable thresholds for issues such as robustness, bias, human oversight, training-data records, and ongoing compliance checks. In tests on two enterprise AI deployments, including a high-risk advisory chatbot, the framework reportedly matched a manual expert audit, found three serious violations, and reduced audit work by about 75%.
New Framework Aims to Standardize Trust Checks for Advanced AI Systems
A new research paper proposes a unified way to evaluate the trustworthiness of large language models, AI agents, and multimodal AI systems. It says benchmark scores alone are not enough, because these systems need different checks for outputs, multi-step actions, tool use, and cross-modal consistency. The framework uses eight areas: capability, robustness, safety, fairness, transparency, governance, oversight, and efficiency. It also calls for uncertainty reporting, traceable evidence, safety-critical overrides, and checks on whether the evaluation process itself is valid and reproducible. The paper links the approach to AI governance needs, including ISO standards, risk management frameworks, and EU AI Act requirements, while noting that real-world validation is still needed.
Study Finds Gaps in U.S. Federal AI Governance Across Vulnerable Sectors
A new study finds that U.S. federal AI governance does not fully match where experts see the biggest sector risks. Researchers reviewed 684 federal AI governance documents across 14 sectors and 24 AI risk areas, measuring both how often risks were mentioned and how deeply they were discussed. The analysis found stronger attention to robustness, system security, governance, public administration, national security, information, and scientific services. By contrast, areas such as finance and healthcare, which experts rated as highly vulnerable, received comparatively less coverage, while socioeconomic, environmental, and newer risks such as multi-agent AI risks were also less addressed.
New Survey Maps Security Risks and Defenses for Agentic AI Systems
A new survey on arXiv examines the security risks of “agentic AI” systems, which can plan tasks, use tools, access memory, run code, and work with other agents. The paper warns that these systems blur the line between normal text input and executable instructions, making attacks such as prompt injection, poisoned memory, data leaks, and unsafe tool use more dangerous. It reviews 206 studies and standards, and proposes a zero-trust security approach using access controls, sandboxing, monitoring, signed embeddings, and checks between AI components. The survey also links these technical safeguards to frameworks such as the NIST AI Risk Management Framework, the EU AI Act, and ISO/IEC 42001, while noting that reliable testing and formal verification remain major open challenges.
AI Governance Framework Aims to Close Enterprise Audit Gaps in Real Time
A September 2026 research paper argues that companies are adopting AI faster than they can govern it, creating an “attestation deficit,” where policies exist but firms cannot prove they are being enforced in time for audits or regulators. Citing reports from Stanford, IBM/Ponemon and EY/AIUC-1, the paper links rising AI incidents, costly breaches, shadow AI and weak monitoring to gaps in governance systems rather than a lack of written rules. It proposes AGIL, a five-layer architecture meant to detect unapproved AI use, score risks, enforce policies in real time, create tamper-evident audit records and update rules across jurisdictions. The framework is still theoretical, with real-world testing left for future work.
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.
