Kia ora! Welcome to New Zealand’s weekly roundup of AI news.
I’m shocked how little NZ coverage I’ve seen beyond the PHP website warnings. The nickname “Black October” reflects the number and seriousness of security flaws and breaches revealed this month.
The wider picture includes banks struggling to distinguish people from agents that browse and click like humans, precautionary government shutdowns in Japan and the Netherlands, exposed documents from Facebook-owner Meta’s Muse, and compromised customer credentials at hosting provider Railway.
These are separate incidents, not all AI-driven. The website warnings concern PHP, software running millions of sites. AI didn’t create these longstanding PHP flaws. It helped researchers discover them and prepare fixes, while making similar work easier for attackers.
Cloudflare saw thousands of PHP attack attempts before disclosure, tens of thousands within 45 minutes afterwards, hundreds of thousands within two hours, and millions by the next day. It keeps updating its protections as attackers adapt.
Want help moving from using AI as a chat bot to delegating tasks? I’m now running:
A four-week 1:1 AI Coaching Programme works around your tools and actual workflows, and
1:1 AI Executive Coaching Programme is open for C-suite leaders in larger companies.
Reply for the details.
Happy reading ✌️
Did someone forward this? Sign up.
🇳🇿 New Zealand News
Four in ten NZ chief executives expect AI to reduce their headcount over the next three years. The NZ Herald’s Mood of the Boardroom survey found 41% expect a decrease, 49% expect no material change and 7% expect an increase, while 54% are already using AI to reduce costs.
Read
Our take: Of the chief executives forecasting lower headcount, 62% already use AI agents. But cost-cutting is the least interesting thing this technology does, and leading with it reflects weak journalism and a lack of ambition about growth. AsureQuality made every AI investment on the survey’s list, including redesigning roles, yet its leadership expects no material change in headcount. Boards should ask which work has improved and what new capabilities AI could offer before deciding the headcount impact. If the only answer is “we have fewer people”, that’s a business with very little ambition. I’d be running for the door if I worked there.
CPA Australia reports widespread AI use but limited integration across Australia and New Zealand, both dead last across the six Asia-Pacific markets surveyed. Its survey of accounting and finance professionals finds 88% use AI and 11% have deeply embedded it; the two countries are reported together.
Read
Our take: Reading this alongside Mood of the Boardroom above, I see the same lack of ambition. Too little effort and strategic thinking goes into changing how the business works and what it could do next. Coming last for deep AI integration across the six surveyed markets should prompt a pretty honest reflection from the ANZ industry. We’re quick to blame politicians, overseas markets and bad luck for our economic problems, but business leaders also need to take responsibility for the decisions within their control, and the opportunities that technologies like AI offer.
Xero’s James Bergin recommends documenting work before choosing AI tools. His new commentary explains how process maps reveal repeated work, missing information and steps that may no longer be needed.
Read
Our take: James’s advice to understand the work and treat AI experiments like school science experiments is sound, with a clear idea of what should improve and a way to measure it. The practical caution is that many small businesses lack the resources to map everything, and technology, staff, products and services can change before that exercise is finished. I’d apply his advice like building with Lego, one brick at a time, documenting one process, testing an improvement and using what was learnt for the next.
Our take: When approved tools weren't overly effective, an investigator used Google Translate to get work done, exposing sensitive material. That’s shadow AI, where staff use unapproved tools because the available options don’t meet their needs. A policy alone won’t fix it. Leaders need usable approved tools, a quick way to raise problems, and checks that the rules work under pressure.
Labour’s Reuben Davidson says the Government’s AI strategy isn’t “fit for purpose”. Speaking on RNZ’s Midday Report, Labour’s technology spokesperson said an Office of AI should be “elevated right to the Cabinet table with a single minister responsible”, with clear rules for data centres and an independent online safety regulator.
Read
Our take: I agree the 2025 strategy has run its course. Being “fast followers of good practice” doesn’t work when the technology changes every month. But an Office of AI set up using existing resources, as Labour’s plan describes, is a coordination unit, not a capability plan. I want Labour to say what it would actually fund: research, talent, and help for businesses stuck in pilots. We’d still be passively watching from the sidelines.
Earth Sciences NZ used generative AI to produce high-resolution climate projections in six hours, work it says would have taken years conventionally. The 12 km dataset covers more than 15,000 simulated years and projects 100-year daily rainfall extremes increasing about 9% per degree of warming.
Read
Our take: This is the kind of AI story NZ should tell more often. The team didn’t trust the model blindly, they’ve trained it on physics-based simulations of a warmer future, then tested it against simulations it had never seen. Councils and infrastructure owners should be all over this data.
Other NZ stories
AI-generated attack ads are finding an audience in the election campaign. University of Canterbury professor Ekant Veer told RNZ a Taxpayers’ Union ad using AI footage of Chris Hipkins stealing wallets reinforces existing beliefs, and that disclosing AI use “doesn’t actually make a huge difference”. Read
The NCSC is urging New Zealand organisations to patch two PHP vulnerabilities. Its 22 October alert reports active exploitation and recommends checking affected websites and embedded systems for compromise. Read
Victoria University has appointed Aotearoa’s first Chair of Advanced Artificial Intelligence. Professor Prasenjit Mitra, most recently at Carnegie Mellon University Africa and previously at Penn State, takes up the Bob Dykes Chair, with a focus on safe and resource-efficient large language models. Read
New Zealand brands and councils are sharing fake Justin Bieber images as a social media joke. The Herald reports that the trend followed confirmation of the singer’s visit and has left some viewers unsure which pictures are real. Read
NZ AI startup SupaHuman has raised $2.6 million for vocational training software. NBR’s report says Movac led the Series A, with other investors participating; the publicly accessible opening does not establish delivery results. Read
NZ conservation groups are using AI to identify pests and monitor native wildlife. The Detail reports on Hark audio devices and AI cameras, while local groups describe the difficulty of funding equipment costing about $1,000 a unit. Read
The idea of working without Wispr Flow feels archaic. For over a year now I've been speaking prompts wherever possible, powering through work at least 3-5x faster. Give it a whirl.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
📚️ Mike's Takes
Helping leaders and teams adapt, learn, and scale with AI.
1. Imagine working at Accenture this past year… Encouraged to use Claude, then questioned over the bill. The sequence raises an awkward question for anyone setting AI budgets.
Read
2. Is your business actually ready for AI agents to join your marketing team? Agents joining the team sounds straightforward until a campaign has to move between people, systems and approvals. How ready is the business for what that requires?
Read
3. My Thursday night, 188 AI agent runs before my allowance resets. A few hours left on the allowance and two projects to get stuck into. What could be finished overnight, and what else could a business delegate this way?
Read
4. How to actually apply AI: The demos look impressive, but how does that translate into something useful in the working day? Six practical steps for making that connection.
Read
5. The race to have your own personalised AI agent continues to heat up. Anthropic has joined the race with Claude Duet. What happens when the agent that knows someone’s personal life starts working alongside them too?
Watch
🎙️ The AI Corner Podcast
How Sharesies Puts AI to Work: Giving Agents Jobs, Keeping Humans in Charge
This week's guest is Nic Cox, Principal Data Engineer at Sharesies. Hear:
Nic explains how existing data models helped Sharesies introduce conversational tools while keeping access permissions and human review in place.
A Claude agent investigates overnight pipeline failures, records findings in Slack and prepares proposed fixes for Nic to review.
Nic describes how shared team knowledge, reusable workflows and usage budgets support AI work across the data team.
🌍 Tech Updates From Global
Recent releases from the companies building AI.
Anthropic
Claude Duet opens a US Pro/Max research preview via waitlist, with morning briefs and approval before messages, payments, signatures or deletions. (Oct 21) Read
Claude Platform introduces breaking API/SDK changes; SDK and CLI licensing switches to Apache 2.0. (Oct 21) Read
Claude Enterprise inference hooks leave beta for customised checks around model inference. (Oct 20) Read
Claude Platform administrators can disable web fetching organisation-wide or restrict which domains agents fetch. (Oct 20) Read
Claude Tag API enters beta for Slack channel settings and access bundles. (Oct 20) Read
Claude Code adds guided auto-mode setup, a read-only Managed Agents security check and broader cloud-provider support. (Oct 19) Read
Claude Code enterprise roles add cloud-session and Remote Control access controls. (Oct 23) Read
OpenAI
ChatGPT Ads reaches 89 countries, adding 26, with ads restricted to Free/Go plans and over 100,000 advertisers. (Oct 21) Read
Decisions API becomes generally available for fast application routing, classification and scoring. (Oct 22) Read
ChatGPT locked chats hide conversations from history and search, requiring an account PIN or supported device authentication. (Oct 20) Read
Codex CLI accepts inline model, reasoning, speed and permission settings, avoiding configuration changes between tasks. (Oct 23) Read
Patch the Planet coordinates fixes for critical PHP vulnerabilities found with frontier AI, alongside US$1 billion in subsidised Daybreak access for defenders. (Oct 21) Read
Google Meet speech translation reaches 63 languages, with automatic detection and per-participant controls on eligible Workspace/Google AI plans. (Oct 22) Read
Google Chat adds recaps with decisions, assigned actions and note links when Meet’s note-taking is enabled. (Oct 19) Read
Gemini reaches Sheets on iOS, enabling eligible users to query data, analyse it and create charts. (Oct 19) Read
Gemini in Connected Sheets analyses complete BigQuery datasets and creates connected charts, pivot tables and dashboards from natural-language instructions. (Oct 19) Read
Gemini Notebook connects inside Gemini for notebook management, source additions and chat queries. (Oct 23) Read
Chinese Frontier Labs
Moonshot releases Kimi K3.1 Pro with 1-million-token context, reporting 65.7% on Terminal-Bench 4.0 and promising full weights by 4 November. (Oct 23) Read
Qwen Code child agents inspect snapshot-isolated workspaces without modifying files, running commands or affecting the parent session. (Oct 19) Read
US Homeland Security orders DeepSeek products, affiliates and third-party implementations excluded from federal civilian procurement and removed from civilian agency systems. (Oct 20) Read
Microsoft
GitHub’s Teams app adds Copilot, letting teams turn issues into fixes from conversations. (Oct 23) Read
GitHub’s Copilot usage API now identifies an organisation’s custom skills, agents and MCP servers. (Oct 22) Read
Copilot’s impact dashboard adds industry comparisons against aggregated agentic-adoption benchmarks. (Oct 22) Read
Copilot model policies add generally available REST endpoints for enterprise, team, agent-app and organisation model-access management. (Oct 19) Read
Copilot Code adds AI coding to Microsoft Copilot. (Oct 21) Read
Copilot Cowork adds federated MCP connectors for live external data within existing user permissions. (Oct 21) Read
Copilot Cowork adds sales, service, education and legal plugins with workflows in Roles & Industry library. (Oct 21) Read
Amazon
Bedrock Advanced Prompt Optimization adds prompt shortening, application inference profiles, and latency/time-to-first-token reporting alongside evaluations. (Oct 20) Read
Bedrock agentic retrieval adds reviewed, versioned instructions for searching managed knowledge bases and answering users. (Oct 19) Read
NVIDIA
NVIDIA’s Minerva combines infrastructure and AI-workload telemetry to detect problems and improve AI factory reliability. (Oct 22) Read
Cosmos Data Factory adds five training-data generation and augmentation blueprints, covering driving scenarios, industrial defects and rare events. (Oct 21) Read
Metropolis video search and summarisation 3.3 creates video AI applications from prompts for monitoring and safety detection. (Oct 21) Read
DGX Spark 64GB systems start rolling out through partners for local models. (Oct 23) Read
Meta
Meta pauses Muse artifact sharing after Claroc discovers over 6,800 shared links through public DNS records. (Oct 20) Read
Meta restores Muse artifact sharing after fixing discoverable URLs and replacing shared links; the flaw affected artifacts, not chats. (Oct 22) Read
Muse reaches 10 million US downloads in 40 days, TechCrunch reports. (Oct 20) Read
Meta completes Prometheus’s first phase and expands the Ohio AI supercluster towards approximately one million GPUs. (Oct 19) Read
Reports & Data This Week
Microsoft finds 84% of UK business leaders expect AI leadership within five years, versus 41% of employees. (Oct 20) Read
Accenture finds 75% of C-suite leaders plan higher AI investment, surveying 3,000 executives and 3,000 employees in 21 countries. (Oct 19) Read
Gartner’s global survey of 203 CEOs/senior executives finds 44% struggle to demonstrate AI financial returns; fieldwork ran in April–May. (Oct 20) Read
Only 13% of 2,002 CIOs/technology executives report significant or transformational AI value in Gartner’s May–July survey. (Oct 19) Read
Forrester predicts one-third of large Asia Pacific enterprises will discover at least ten unauthorised employee-built agents per approved agent in 2027. (Oct 22) Read
Forrester predicts AI will consume 20% of business technology spending in 2027, increasing cost and infrastructure scrutiny. (Oct 21) Read
Forrester predicts AI agents will influence 75% of business purchasing decisions in 2027 as buyers seek human validation. (Oct 21) Read
👋 Mike & Erin