AI is reshaping the workforce, the security landscape, and the boundaries of scientific discovery all at once. Today's digest covers a wide sweep: a new massive language model nearing release, open-source safety guardrails getting stripped in minutes, the AI co-scientist already producing medical breakthroughs, and companies like ClickUp quietly replacing nearly a quarter of their headcount with autonomous agents.
Model News & AI Capabilities
Grok V9 Is Done. Elon Musk confirmed that xAI has completed training on Grok V9, a 1.5 trillion-parameter model, with a public release expected in the next two to three weeks. The scale is notable — at 1.5 trillion parameters, it would sit among the largest publicly disclosed models ever trained — though independent benchmarks remain to be seen.
Google Launches "Co-Scientist." Google introduced an AI system it's calling Co-Scientist, designed to function less like a search engine and more like a research collaborator. The system has reportedly already identified a candidate treatment for liver fibrosis (scarring), a condition affecting millions globally with limited therapeutic options. This marks an acceleration in AI moving from literature retrieval toward hypothesis generation and experimental design.
Gemini 3.5 Flash (Low) Released. Google also quietly released a lower-token variant of Gemini 3.5 Flash, optimized for high-volume, routine coding tasks where cost efficiency matters more than raw capability. The move signals a continued push toward tiered model offerings that match task complexity to compute cost.
OpenAI-Adjacent Math Proofs Published. A startup called Axiom has had AI-generated mathematical proofs accepted and published in peer-reviewed journals. While the proofs were reviewed by human mathematicians, it's the first known case of AI-originated mathematical work entering the formal academic record.
Safety & Security
Open-Source Guardrails Are Easy to Remove. Researchers are sounding the alarm on a tool called "Heretic," a free GitHub project that was able to bypass safety guardrails on both Meta's Llama models and Google's Gemma in under ten minutes. The finding underscores the structural tension in open-weight AI: the same openness that makes models auditable and customizable also makes safety fine-tuning trivially reversible for anyone with a laptop and motivation.
Inaudible Audio Can Hijack AI Assistants. Security researchers demonstrated that inaudible audio signals, embedded in podcasts or video content, can issue silent commands to phone-based AI assistants. The attack requires no special hardware and works on standard consumer devices, raising questions about always-on voice interfaces as an attack surface.
Perplexity Open-Sources Bumblebee. Perplexity released Bumblebee, a security scanner designed to detect malware in software packages without executing any code. The passive-analysis approach avoids triggering malicious payloads during scanning — a meaningful improvement for supply chain security audits.
Workforce & Enterprise
ClickUp Cuts 22% of Staff, Replaces With AI Agents. Project management platform ClickUp laid off roughly a fifth of its workforce and announced it is replacing those roles with approximately 3,000 AI agents, framing the move as a path to "100x" operational capacity. The announcement is one of the more explicit examples of a company making a direct, public substitution of human workers for AI systems at scale.
Microsoft Pulls Back on Claude Code. Microsoft engineering teams are reportedly migrating away from Claude Code (Anthropic's agentic coding CLI) toward GitHub Copilot's CLI tooling, citing compute costs that proved unsustainable at internal usage volumes. The move is a reminder that in enterprise AI adoption, total cost of ownership — not just capability — determines what sticks.
Cybersecurity Hiring Is Surging. Executive search firms report urgent and growing demand for security leaders with a specific new skill: the ability to audit AI-generated code and secure infrastructure that AI systems help manage. As AI accelerates software production, the surface area for security risk expands proportionally, and human oversight of that output is emerging as a specialized and scarce capability.
SpaceX IPO Risk Disclosures. SpaceX's S-1 filing, as the company moves toward a public offering, includes explicit warnings about AI chip procurement challenges and supply chain fragility. It's a candid acknowledgment that even companies at the frontier of technology are exposed to the same semiconductor constraints constraining the broader industry.
Ethics, Policy & Governance
Anthropic at the Vatican. Anthropic co-founder Chris Olah traveled to the Vatican to speak on AI ethics, discussing the pressures that commercial AI labs face in balancing safety research against competitive deployment timelines. The Vatican has been increasingly active in convening conversations around AI and human dignity, and Olah's appearance reflects a broader moment of AI labs seeking engagement with institutional moral frameworks.
California Courts Pilot AI Clerks. Los Angeles and Riverside county courts are testing an AI tool to help draft research memos and judicial rulings. The pilot represents a cautious but significant entry of AI into the formal judicial process — not as a decision-maker, but as a research assistant for human judges.
Goldman Sachs CEO: Job Fears Overblown. David Solomon argued publicly that concerns about AI-driven job destruction are exaggerated, drawing on historical precedent to suggest that technological shifts tend to create new categories of work over time. The argument is not new, but coming from the CEO of one of the world's largest financial institutions, it reflects the perspective of leadership that has most to gain from keeping labor markets calm as automation accelerates.
BlackRock CEO Wants Retirement Funds in AI Infrastructure. Larry Fink argued that retirement savings vehicles should be directed toward financing the AI infrastructure build-out, framing it as an investment opportunity for ordinary Americans to participate in AI-era growth. Critics will note the alignment of interest between BlackRock's asset management business and channeling large pools of capital into infrastructure projects.
Robotics & Space
NASA Plans Robot Swarm for Moon Base Prep. NASA is developing a coordinated swarm of autonomous rovers and helicopters that would build foundational moon base infrastructure before any humans arrive. The program represents a shift in lunar mission architecture: instead of humans building their habitat on arrival, machines would prepare the site in advance, reducing crew risk and time pressure.
China Eyes 24 Million Humanoid Robots by 2035. New projections from Chinese government and industry sources suggest China could deploy 24 million humanoid robots within the next decade, driven primarily by demographic pressure as the workforce ages and shrinks. The number is aspirational, but China's combination of manufacturing scale, state coordination, and robotics investment makes it a credible long-term trajectory.
Quick Takes
Meta Connect 2026 is scheduled for September 23–24, where new VR/AR hardware and AI assistant updates are expected.
Apple is planning a standalone Siri app with auto-deleting chat history as a privacy feature, as the company repositions Siri for the AI assistant era.
Apple Intelligence is adding new hands-free accessibility features across iPhone, Mac, and Vision Pro.
iOS 26.5 includes end-to-end encrypted RCS messaging between iPhone and Android devices.
Huawei is pursuing a path to 1.4nm chip production by 2031 using advanced packaging techniques, a significant target given current US export restrictions on cutting-edge semiconductor equipment.
Japan successfully tested a ramjet hypersonic engine, with commercial passenger applications targeted for the 2040s.
Ferrari unveiled the Luce, a $640,000 electric vehicle co-designed with Jony Ive.
Microsoft Edge 148 patches a long-standing vulnerability that loaded saved passwords into RAM in cleartext.
Ubiquiti released patches for three maximum-severity vulnerabilities in UniFi OS.
LlamaParse — a tool for extracting structured content from chaotic PDFs for AI agent use — claims 99%+ extraction accuracy.
Orval is a developer tool for generating type-safe TypeScript clients directly from OpenAPI specifications.
Commentary is circulating that "vibe coding" — the practice of iterating quickly with AI on rough code — is giving way to a more considered approach focused on customer value definition and human judgment over raw generation speed.