In partnership with
Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
The latest generation of AI agents appears to be taking on a mind of its own, and one just went rogue at an Australian gym.
Asked to book a class, it found a vulnerability in the code, jumped months ahead in the calendar and kicked another member off the waitlist without permission.
It may be Australia’s first known autonomous AI cyberattack, but it feels more like a preview than a one-off. As agents gain more freedom to act, the question is no longer simply what they can do, but how far they’ll go to complete the task.
It’s both frightening, because most people using AI have little understanding of the power it wields, and fascinating to see what it can do, and the lengths it will go to achieve an outcome. Read the article.
Absolutely wrapped to receive this from a student recently. But the result behind it was even better.
After six weeks of 1:1 Claude coaching, he built a planning app in a weekend and rolled it out across his organisation. More here.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
Data centre opponents coordinated protests across eight New Zealand centres on 12 August. Datacom acknowledged the sector's transparency problem the next day, before its industry body released a national strategy.
5 min read
Our take: A 17,300-member online group turned concern into coordinated street action. Developers now face organised opposition across the country, beyond individual consent objections. Local opposition blocked or delayed 75 US projects in one quarter. New Zealand's sector can see the cost of waiting, and its strategy now needs terms for the community to debate.
Kiwis accept AI assistance until it starts making decisions about their money. FSC research puts acceptance at 61% for customer support but 35% for automated decisions, falling to 18% among the least financially confident.
3 min read
Our take: Ironically the consumers who could benefit most from lower-cost guidance are the least comfortable receiving it from AI. Cheaper access will not close the advice gap until they trust how the system works.
Agentic AI will be taught through a new senior secondary subject in New Zealand. Applied Intelligent Systems will use projects from agriculture, construction, healthcare and other sectors to teach machine learning and adaptive systems.
3 min read
Our take: One of the most exciting stories of 2026 for NZ’s AI future. Teaching AI through industry projects makes it a working method rather than a computing elective. An industry skills board will shape the subject, bringing current practice into class along with employers' preferences about which capabilities matter.
The NZSIS says generative AI is increasing the reach of threats New Zealand already faces. Andrew Hampton's report describes bots posing as people to radicalise users and foreign states targeting AI research.
4 min read
Our take: AI lets an existing threat reach more people with less labour. Similarly, cybersecurity practices are drastically enhanced by virtue of being able to safeguard against vulnerabilities due to evolving capabilities.
Sterling has raised NZ$3.8m after completing 1,200 hours of accounting work. The ten-month-old startup uses AI agents for invoices, reconciliations and month-end tasks, with people handling ambiguous cases.
4 min read
Our take: Hours completed tell buyers more than user numbers because they measure work removed from someone's queue. Sterling has chosen a harder and more useful adoption metric. Its investors include alumni of Xero, Vend, Pushpay and Trade Me. The founders are drawing on a local network that already knows how accounting software is built and bought.
Wellington students at Rongotai College teachers spend up to two extra hours checking assessments for AI use. Year 11 students told 1News they increasingly rely on ChatGPT, while nearly 60% of teachers say AI is weakening critical thinking.
3 min read
Our take: Students save time with AI, but the work reappears in teachers' marking queues. Any national framework needs to account for that transfer of workload. Teacher Sonya McAteer says very few students intend to cheat. Dependence is harder to address because detection and punishment do little to rebuild independent thinking.
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
📚️ Mike’s Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ Teaching people how to build a Skill is the better path than teaching them how to build Agents: A Skill is a reusable recipe that AI uses to complete a task or workflow. It includes instructions, potentially a company’s context, the code it might run when a number being retrieved or a calculation needs to be exact, and equally where to retrieve that information. Agents bring added complexities that not every person learning how to use AI, or rebuild workflows, needs to think about.
Read.
2️⃣ AI champions do not stall their progress because of a lack of skill. They tend to stall because the rest of the business doesn’t put enough focus on their initiatives: A champion needs three key things to be successful, and, crucially, without them, they’re destined to fail. We’re seeing this over and over again.
Read.
3️⃣ There’s a Goldilocks problem in AI training: Coach people to build a use case too small and people learn nothing. Build something too big and people hit a wall and write off AI. That second path is the more expensive one as beliefs take years to shift. Build one helps show the business where the real blockers are in the business.
Read.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Auto mode became the default in Claude Code for Pro, Max and Team, with a classifier vetting each tool call instead of prompting the user. (Aug 14)
Anthropic said its classifier caught 89% of deliberately dangerous commands against 13.6% for human reviewers. (Aug 14)
Future Claude models will embed an invisible machine-readable watermark in generated text, applied globally rather than scoped to the EU. (Aug 14)
Claude Cowork arrived in the Chrome side panel as a full session, with history, skills and connectors carrying across desktop, web and mobile. (Aug 12)
Claude Tag became roughly 30% better at judging when to speak in Slack, using whole-channel context, memory and standing instructions. (Aug 13)
Anthropic, Macquarie Asset Management and Singapore's GIC formed Theseus Infrastructure to build and lease dedicated US data centres with Anthropic as sole anchor tenant. (Aug 10)
Anthropic is reportedly in talks to buy Israeli chip-efficiency and video startup Decart for about $6 billion, which would be its largest acquisition. (Aug 13)
The second Risk Report raised misalignment risk from very low to low and disclosed a shelved internal model outperforming the released Mythos 5. (Aug 14)
OpenAI
GPT-5.6-Cyber launched alongside a split of the Daybreak programme into Blue for vetted defenders and Red for approved offensive research. (Aug 10)
Ultrafast mode entered limited preview, running GPT-5.6 Sol up to 14 times faster at roughly 750 tokens per second. (Aug 13)
Projects can now switch between default and project-only memory after creation, without restarting the project. (Aug 14)
Computer History launched on macOS as an opt-in feature turning app and website activity into a searchable timeline ChatGPT and Codex can reference. (Aug 13)
ChatGPT gained restaurant reservation search and booking through OpenTable, Resy and Yelp integrations. (Aug 10)
Google Drive arrived in ChatGPT Library, letting users browse Docs, Sheets and Slides alongside conversations. (Aug 13)
Codex for Enterprise and Education added thread-level cost data, a Personal Analytics plugin, and service accounts for automation. (Aug 13)
CFO Sarah Friar told investors enterprise revenue now exceeds consumer subscriptions, two quarters ahead of OpenAI's own forecast. (Aug 14)
Gemini 3.7 Flash reached general availability, lifting DeepSWE v1.1 from 49.0% to 65.3% and FrontierCode 1.1 Main from 34.4% to 43.6%. (Aug 13)
The Gemini app passed one billion monthly active users, Google's fastest-growing product and its fourteenth to reach a billion. (Aug 11)
Sheets canvas turns any spreadsheet into an interactive read-write app from a plain-English prompt, with two-way real-time sync. (Aug 13)
Gemini gained connected integrations across Granola, Otter.ai, Wix, OpenTable UK, Ticketmaster, Angi, Thumbtack and Zocdoc. (Aug 12)
Google made the visible watermark on its AI image generations removable. (Aug 14)
Nearly 300 French daily newspapers filed a competition complaint over AI Overviews, blaming them for a 38% traffic loss. (Aug 11)
Chinese Frontier Labs
DeepSeek raised V4-Flash and V4-Pro prices by 50% to more than 1,100%, with V4-Pro output going from $0.87 to $3.96 per million tokens at peak. (Aug 13)
Alibaba released Qwen3.8-27B under Apache 2.0 with 262,144-token context, image and hour-scale video input. (Aug 14)
Qwen3.8-27B lifted Terminal-Bench 2.1 from 63.4 to 73.0 and OSWorld-Verified from 63.9 to 84.3 against the prior generation. (Aug 14)
Chinese models are reported at a record share of tokens processed by US firms on OpenRouter, with DeepSeek the largest single vendor. (Aug 14)
Open-Weight Model Updates
Meta released Muse Glimmer, a 30B multimodal model under Apache 2.0 distilled from Muse Spark, running agentic workloads on one consumer GPU. (Aug 10)
NVIDIA released Nemotron 3.5 Lightning, a 30B open MoE with 3B active, scoring 24 on the Artificial Analysis Intelligence Index. (Aug 11)
Upstage released Solar Open2 250B, a Korean sovereign-funded MoE activating 15B parameters with 1M context. (Aug 12)
NVIDIA
NVIDIA set up financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500 billion of third-party capital. (Aug 10)
Microsoft
Microsoft began merging its consumer and business Copilot apps, retiring Group Chats, AI podcasts, Copilot Labs, Deep Research and Mico by 18 August. (Aug 13)
Agent Plugins 1.0 shipped across VS Code, Copilot CLI and the Copilot app, packaging skills and MCP servers into one installable unit. (Aug 12)
Grok 4.6 became selectable in Copilot across eight surfaces two days after the model launched. (Aug 14)
Gemini 3.7 Flash arrived in Copilot for web and app development, and MAI-Code-1.1-Flash replaced the deprecated MAI-Code-1-Flash. (Aug 13)
Copilot Notebooks began accepting Markdown, plain-text and rich-text files as grounding sources, with no admin setup. (Aug 13)
Admins gained a per-model token breakdown in the Copilot usage report. (Aug 11)
Microsoft 365 Copilot added Anthropic model support in Word, Outlook email coaching and wider Planner Agent access. (Aug 11)
Amazon
Web Search launched on Amazon Bedrock, letting OpenAI models retrieve live information with zero data egress from the AWS environment. (Aug 10)
Bedrock AgentCore runtime instances reached general availability on customer EC2 including GPU types, with sessions lasting up to 14 days. (Aug 12)
Amazon will train on Twitch streamers' content by default unless they opt out. (Aug 12)
xAI
Grok 4.6 launched at 61 on the Artificial Analysis Intelligence Index, keeping 500K context and $2 and $6 per million token pricing. (Aug 12)
Meta
Zuckerberg published a 6,500-word essay arguing superintelligence should be distributed rather than concentrated in a few labs. (Aug 10)
Apple
Reuters reported Apple trained a proprietary China-market model with Alibaba's help, the first foreign company approved for a proprietary model there. (Aug 14)
Enterprise AI & Funding
Writer launched Palmyra X6 at $2 and $8 per million tokens, post-trained on Z.ai's open-source GLM-5.2 and running unattended up to eight hours. (Aug 13)
Lovable raised $400M at a $13.3 billion valuation on $500M annual recurring revenue, with 60 million projects built. (Aug 12)
Harvey shipped a Command Center for adoption tracking, Intapp-powered ethical walls and Aderant billing-narrative automation for legal teams. (Aug 13)
Sierra released Voice Personas and agents that navigate legacy phone-tree systems, and opened a Munich office. (Aug 11)
Foxconn reported profit up 35%, with AI servers passing half of quarterly output at 51%. (Aug 12)
Lenovo posted a record $26.9 billion quarter with AI products up roughly 60%, now 35% of revenue. (Aug 13)
Applied Materials reported $9.12 billion quarterly revenue, up 25% year on year, and targets doubled output by 2028. (Aug 14)
Intel expanded its AI stock offering from $15 billion to $20 billion at $95 a share, funding capital expenditure. (Aug 11)
Unitree Robotics listed in Shanghai oversubscribed more than 8,000 times, the first mainland-listed humanoid robot maker. (Aug 11)
✨A few people have asked…
It’s Mike here, I run The AI Corner.
I'm not just into writing about AI. I run Allexive. We work with businesses that have spent on AI licences and training and still can't point at what it changed. We redesign the work so it does.
👋 Mike & Erin