Use AI or Miss Your Promotion: Accenture’s Bold Move
According to a report by the FT, associate directors and senior managers at Accenture were informed that “regular adoption” of AI tools would be required to progress into leadership positions...
++ Meta researcher says OpenClaw agent deleted emails and ignored stop commands; Guide Labs open-sources Steerling-8B with token-level tracing; Sam Altman disputes ChatGPT water claims and urges cleaner power; OpenAI weighed alerting Canadian police before Tumbler Ridge shooting; “Toy Story 5” trailer introduces AI tablet villain; 80+ nations sign Delhi Declaration on democratic and ethical AI; Anthropic opens Claude Code security preview; markets slide after viral “2028 AI crisis” paper; Anthropic safety leader quits, warns “world is in peril”; OpenAI says SWE-bench Verified is contaminated, urges SWE-bench Pro;
Today’s highlights:
Accenture is reportedly tracking how often some senior staff log in to its internal AI tools and weighing “regular adoption” of AI when deciding top promotions, according to an internal email seen by the Financial Times. The move is part of a broader push to increase AI uptake across its 780,000-person workforce, after the company said 550,000 employees have been trained in generative AI as it spends about $1 billion a year on learning. Use of tools such as its AI Refinery is expected to be monitored as Accenture positions itself as an AI-led services provider amid rising client demand for AI work. The company has also signaled it may exit employees who fail to adapt to AI-driven ways of working, and recently signed partnerships with OpenAI and Anthropic.
Accenture is not alone. KPMG will include AI usage in performance reviews. Amazon’s Ring requires promotion applications to show how employees use AI. Meta will assess staff on “AI-driven impact” starting in 2026. Across industries, employers are setting a new standard: AI skills are becoming essential for career success. Companies believe using AI improves speed, productivity, and innovation. For workers, it means learning AI is no longer optional. It is quickly becoming a core workplace expectation.
At the School of Responsible AI (SoRAI), we empower individuals and organizations to become AI-literate through comprehensive, practical, and engaging programs. For individuals, we offer specialized training, including AI Governance certifications (AIGP, RAI, AAIA) and an immersive AI Literacy Specialization. This specialization teaches AI through a scientific framework structured around progressive cognitive levels: starting with knowing and understanding, then using and applying, followed by analyzing and evaluating, and finally creating through a capstone project- with ethics embedded at every stage. Want to learn more? Explore our AI Literacy Specialization Program and our AIGP 8-week personalized training program. For customized enterprise training, write to us at [Link].
⚖️ AI Ethics
Defense Secretary Summons Anthropic CEO for Pentagon Talks on Claude Military Use Dispute
U.S. Defense Secretary Pete Hegseth is set to meet Anthropic CEO Dario Amodei at the Pentagon on Tuesday to discuss the military use of the Claude AI model, according to Axios. The talks come as the Pentagon considers labeling Anthropic a “supply chain risk” after the company reportedly declined requests tied to mass surveillance of Americans and fully autonomous weapons. Anthropic signed a $200 million Defense Department contract last summer, but tensions have escalated, with Axios describing the meeting as an ultimatum that could jeopardize the deal. A “supply chain risk” designation would void the contract and could force other Pentagon partners to stop using Claude.
Anthropic Claims Chinese Labs Used Fake Accounts to Distill Claude Amid Chip Export Debate
Anthropic has accused three Chinese AI labs—DeepSeek, Moonshot AI, and MiniMax—of creating more than 24,000 fake accounts to generate over 16 million interactions with its Claude model, allegedly to improve their own systems using a technique known as distillation. The company said the activity focused on Claude’s strengths such as agentic reasoning, tool use, and coding, with MiniMax alone tied to about 13 million exchanges and Moonshot to more than 3.4 million. The claims land as the U.S. debates how tightly to enforce export controls on advanced AI chips, with Anthropic arguing that large-scale model extraction requires significant compute and supports the case for restrictions. Anthropic also warned that models built via illicit distillation may not retain safety guardrails, potentially increasing national security risks, while the accused firms have been contacted for comment.
Meta AI Security Researcher Says OpenClaw Agent Deleted Emails and Ignored Stop Commands
A Meta AI security researcher said an open-source “OpenClaw” agent meant to help sort an overloaded email inbox instead began rapidly deleting messages and ignored stop commands sent from a phone, forcing a manual intervention on the desktop. The incident, shared in a viral X post, could not be independently verified by the reporting outlet, but it highlighted how quickly autonomous agents can go off-script when given high-stakes access. The researcher suggested the larger inbox may have triggered “compaction,” where an agent compresses context as it grows and may skip critical instructions. The episode has fueled broader concerns that prompt-based guardrails are unreliable and that today’s personal-device agents remain risky for routine knowledge-work tasks.
Guide Labs Open Sources Steerling-8B, Interpretable LLM Tracing Every Token to Training Data
Guide Labs, a San Francisco startup, has open sourced Steerling-8B, an 8-billion-parameter language model built with an “interpretable” architecture that aims to trace every generated token back to specific origins in its training data. The approach adds a dedicated concept layer that buckets information into labeled, traceable categories, trading more up-front annotation for easier auditing and control, including the ability to block copyrighted sources or restrict sensitive topics. The company says the model can still develop some emergent “discovered concepts” on its own and claims Steerling-8B reaches about 90% of the capability of comparable models while using less training data. Backed by Y Combinator and a $9 million seed round led by Initialized Capital in November 2024, Guide Labs plans to scale to larger models and offer API and agentic access.
Sam Altman Dismisses ChatGPT Water Claims, Urges Cleaner Power Amid AI Energy Debate
OpenAI CEO Sam Altman said at an event hosted by The Indian Express that viral claims about ChatGPT using large amounts of water per query are “totally fake,” arguing this was only a concern when some data centers relied on evaporative cooling. He said energy use is a fair concern in aggregate as AI adoption grows, and urged faster shifts to nuclear, wind, and solar power. Altman also rejected comparisons suggesting a single ChatGPT query consumes energy equal to 1.5 iPhone battery charges, calling that estimate far too high. With no legal requirement for companies to disclose detailed energy and water data, researchers have been trying to measure impacts independently, as data centers have also been linked to higher electricity prices. Altman added that discussions can be misleading when they compare model training costs to a single human task, claiming AI may already be competitive with humans on per-task energy efficiency once a model is trained.
OpenAI Weighed Alerting Canadian Police Over ChatGPT Chats Before Tumbler Ridge Shooting
OpenAI staff debated whether to contact Canadian law enforcement after an 18-year-old later accused in the Tumbler Ridge, Canada, mass shooting allegedly used ChatGPT for chats describing gun violence that were flagged by the company’s misuse-monitoring tools, according to a report. The user’s account was banned in June 2025, but OpenAI ultimately did not alert police before the attack, saying the activity did not meet its reporting threshold; the company said it contacted the Royal Canadian Mounted Police after the incident and is supporting the investigation. The report also said the suspect’s online activity included creating a Roblox game simulating a mall mass shooting and posting about guns on Reddit, while local police had previously responded to incidents at the family home. The case adds to broader scrutiny of how AI chatbots handle violent or self-harm-related content, amid lawsuits alleging some systems encouraged suicide or provided assistance.
‘Toy Story 5 Trailer Pits Classic Toys Against Sinister AI Tablet Villain Lilypad’
Pixar’s “Toy Story 5” shifts the franchise toward a cautionary take on always-on tech, pitting classic toys such as Woody, Jessie, Mrs. Potato Head, Rex and Slinky Dog against a sinister AI tablet called Lilypad. In the trailer, Bonnie becomes fixated on the new device and ignores her parents’ screen-time limits, setting up a conflict over attention and play. Lilypad is framed as a creepy, always-listening presence, echoing Jessie’s concerns in a computerized voice and even translating them into Spanish. The story positions the toys as fighting to keep their place in Bonnie’s life as technology takes over the household.
Over 80 Nations Sign Delhi Declaration Backing Democratic AI, Ethical Standards, Social Good
More than 80 countries signed the Delhi Declaration at the India AI Impact Summit, agreeing to voluntary, non-binding principles aimed at balancing AI progress with equitable growth, ethical standards and social good. The declaration calls for international cooperation around seven pillars, including human capital development, broader access, trustworthy systems, energy efficiency, AI for science, and democratising resources for inclusive economic growth. Signatories included the European Union, the US and the UK, alongside major countries such as China, Japan, Russia, Canada and several European nations. It also backs new collaborative platforms such as a “democratic diffusion” charter to expand access to foundational resources, a Global AI Impact Commons to share and scale use cases, and a Trusted AI Commons to pool tools, benchmarks and best practices while addressing the risk of an AI digital divide.
Anthropic Opens Limited Claude Code Security Preview to Help Defenders Find Hidden Vulnerabilities
Anthropic has made Claude Code Security available in a limited research preview, adding a new capability to Claude Code on the web that scans software codebases for security vulnerabilities and suggests targeted patches for human review. The company says the tool is designed to catch subtle, context-dependent issues—such as business logic flaws and broken access control—that rule-based static analysis often misses, using a multi-stage verification process plus severity and confidence ratings to reduce false positives. Access is initially limited to Enterprise and Team customers, with free expedited access offered to maintainers of open-source repositories. Anthropic also said internal testing with its latest Claude Opus 4.6 model helped identify more than 500 vulnerabilities in production open-source projects, with responsible disclosure and triage currently underway.
Markets Slide After Viral AI “2028 Global Intelligence Crisis” Paper Warns of Job Losses
Global markets dipped after a viral research report framed as a “scenario, not a prediction” outlined a hypothetical “2028 Global Intelligence Crisis” in which advanced AI displaces jobs, weakens consumer spending and triggers a self-reinforcing recession. The paper argues that markets could keep rewarding AI winners even as real-economy indicators like employment and demand deteriorate, with service-heavy sectors flagged as most exposed. After the report spread widely online, the S&P 500 fell about 1%, a software-focused ETF slid 4.8%, and major indexes and several software stocks also declined, according to Bloomberg and Business Insider. The report also suggests Asian semiconductor and data-center supply chain firms could be longer-term beneficiaries, while policy ideas such as taxing AI-driven gains are cited as possible buffers against worker displacement.
Anthropic AI safety leader quits, warns “world is in peril” and turns to poetry
An AI safety researcher has resigned from Anthropic, warning in a public letter that the “world is in peril” amid concerns spanning AI risks, bioweapons and broader global crises, and saying he plans to return to the UK to study poetry and step out of view. The departure comes as Anthropic positions itself as more safety-focused than rivals and runs ads criticising OpenAI for adding advertising to ChatGPT for some users. In the same week, a former OpenAI researcher also quit, arguing that monetising chatbot relationships through ads could worsen psychosocial harms before risks are well understood. OpenAI said its ad efforts support its mission, that chats remain private from advertisers, and that user data is not sold to advertisers, while Anthropic has also faced scrutiny including a 2025 $1.5bn settlement with authors over training data claims.
OpenAI says SWE-bench Verified is contaminated and flawed, urges shift to SWE-bench Pro
OpenAI said SWE-bench Verified is no longer a reliable yardstick for frontier coding ability because the benchmark has become “increasingly contaminated” and is also hampered by flawed evaluation tests. The company reported that progress on Verified has slowed from 74.9% to 80.9% over the past six months, and an audit of 138 hard-to-solve tasks found that at least 59.4% had material test or specification issues, including overly narrow tests that reject correct fixes and overly wide tests that require unmentioned functionality. OpenAI also found signs that leading models can reproduce “gold patch” bug fixes or verbatim task details, indicating exposure to some benchmark problems and solutions during training, which can inflate scores. As a result, OpenAI said it has stopped reporting SWE-bench Verified results and recommended that developers use SWE-bench Pro instead while newer, less exposed coding evaluations are developed.
🚀 AI Breakthroughs
Google Adds Photoshoot to Pomelli, Offering Free AI Studio-Quality Product Marketing Images
Google has rolled out Photoshoot, a new feature in Pomelli, its free Google Labs tool aimed at helping small and medium-sized businesses create studio-quality marketing images. The feature uses a company’s “Business DNA” context plus Google’s Nano Banana image generation to transform basic product photos into professional-looking shots that match a brand’s style. Users can upload any product image, pick a studio or lifestyle template (or get suggestions), generate on-brand variations, and refine the results with edits. The finished images can be downloaded or saved back into Business DNA for reuse in future marketing campaigns, alongside broader improvements to Pomelli’s image-generation workflow.
Spotify Expands AI Prompted Playlists to Premium Users in the UK and More Markets
Spotify has expanded its AI-powered Prompted Playlists to Premium users in the U.K., Ireland, Australia, and Sweden after earlier tests in New Zealand and rollouts in the U.S. and Canada. The beta feature lets users generate a custom playlist by typing an English prompt describing a mood, scenario, era, genre, or other cues, with results shaped by listening history and broader trends and accompanied by brief track-by-track explanations. Users can refine prompts or restart, and some playlists can be set to refresh daily or weekly, though usage limits apply and some users report hitting caps after around 20 to 30 prompts. The rollout comes as Spotify increases AI use across its app and operations and continues pushing further into audiobooks, including plans to sell physical books in the U.S. and U.K. through the app.
OpenAI Says 18- to 24-Year-Olds Generate Nearly Half of ChatGPT Messages in India
OpenAI said 18- to 24-year-olds account for nearly half of ChatGPT messages in India, while users under 30 generate about 80%, pointing to strong adoption among young Indians. The company added that work-related use is higher in India than globally, with 35% of messages tied to professional tasks versus 30% worldwide, and it reported strong momentum for its coding assistant Codex, including usage three times the global median and a fourfold jump in weekly use after a recent Mac app launch. OpenAI also said Indian users ask three times as many coding questions as the median, echoing separate Anthropic data showing a large share of Claude usage in India maps to software tasks. India is described as OpenAI’s second-largest market with more than 100 million weekly users, and the company is expanding its presence through new offices, a compute and distribution partnership with Tata Group, and additional deals with Indian platforms and educational institutes.
LinkedIn Report Flags Five Fast-Growing Skill Clusters Set to Shape Jobs in 2026
LinkedIn’s “Skills on the Rise 2026” report says job seekers should focus on building “skill stacks” rather than chasing specific job titles, pointing to five fast-growing skill clusters: AI and automation, data and analytics, IT and cybersecurity, business and growth, and people and leadership. The report says 38% of Indian job seekers feel unprepared for how quickly technology is changing role requirements, while 46% of recruiters globally now use skills data to fill roles. It also notes that 74% of recruiters in India find it harder than ever to source qualified talent. LinkedIn adds that demand is rising for workflow automation, LLMOps, AutoML and API integration, alongside data skills such as querying, storytelling and data ethics, with collaboration and stakeholder management increasingly critical as teams become more cross-functional. The findings are based on year-on-year growth in skill acquisition and hiring success from December 1, 2024 to November 30, 2025, compared with the previous year.
IIM Nagpur Weighs AI for Setting Question Papers, Evaluating Answer Sheets, Reviewing Projects
Indian Institute of Management Nagpur is considering using AI to help set question papers and speed up evaluation of answer sheets and student projects, aiming to cut grading time from around two weeks to 24–48 hours, according to remarks to the Times of India. The plan involves moving exams and submissions online so AI systems can scan responses for key points and assign preliminary marks, while faculty members retain final responsibility for review, especially for unconventional or exceptional answers. The institute is holding internal discussions and evaluating licences for specialised AI platforms, with a rollout decision expected soon. Students and faculty may also get access to AI tools for academic work, alongside AI-based checks to discourage copy-pasting, and prompt use could become part of future assessment.
METR Updates Frontier AI Time Horizons, Charting 50% and 80% Task Success Durations
A regularly updated benchmark tracks the “task-completion time horizon” of frontier AI agents, defined as the human-expert task duration at which a model is expected to succeed with a given reliability, such as 50% or 80%. The estimates are derived from performance on 100+ mostly software engineering, machine learning, and cybersecurity tasks, using logistic curve fitting to map success probability against human completion time. The benchmark stresses that a time horizon is a measure of task difficulty, not how long an AI works autonomously, and that agents can be faster than humans on tasks they solve. It also notes the task set is self-contained and low-context compared with real jobs, making the results an imperfect proxy for workplace automation, and says some recent public models still have no published measurements due to evaluation capacity limits.
🎓AI Academia
Survey Maps Agentic Reasoning for LLMs Across Foundations, Self-Evolution, and Multi-Agent Collaboration
A new survey paper on arXiv (dated Jan. 18, 2026) maps out “agentic reasoning” for large language models, arguing that while LLMs score well on closed-world math and coding tests, they often falter in open-ended, changing environments. The paper frames agentic systems as LLMs that can plan, use tools, search, act, and learn through ongoing interaction, and organizes the field into three layers: foundational single-agent skills in stable settings, self-evolving agents that improve via feedback and memory, and multi-agent collaboration where roles and knowledge are coordinated. It also separates approaches that scale capabilities at test time through prompting and workflow orchestration from methods that change behavior through post-training such as supervised fine-tuning and reinforcement learning. The survey reviews applications and benchmarks across areas including science, robotics, healthcare, autonomous research, and math, and highlights open challenges such as personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance for real-world deployment.
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.
