Possibl AI

The most important AI conversation in New Zealand yesterday wasn't about models. It was about trust. I spent the day at the Digital Trust Hui Taumata in Wellington, and the theme "Towards Universal Trust" landed harder than expected. The fireside chat with Drummond Reed and Dr Karaitiana Taiuru JP, PhD, MInstD set the tone: sovereign AI isn't "too hard" it just needs breaking down into solvable pieces, with Te Tiriti and Māori data sovereignty built into the architecture rather than bolted on. As I saw at the round tables, two threads stood out for anyone deploying AI in their business. First, agentic identity is now a live issue: when an AI agent acts in your name, who answers for it? Second, as Joel Foster put it, identity must be proven, not assumed. AI has made deception cheap; the old trust signals no longer hold. For NZ organisations building with agents, identity and governance aren't compliance afterthoughts. They're the foundation. Great to hear from Maria Robertson, Hon Paul Goldsmith, Xavier Russo, Sigurd Magnusson, Helen Littlewood and Dr Pramod Varma. Thanks to Digital Identity NZ for a well-run day and Andy Higgs for another great HUI. What would it take for you to trust an AI agent acting on your behalf? #AIGovernance #AgenticAI #NZTech Digital Identity New Zealand Andy Higgs Possibl AI

1 in 4 public AI agent skills carries a security flaw. 26.1% of 31,132 skills analysed this year, to be exact. A single marketplace audit found 76 confirmed malicious payloads in circulation: credential theft, backdoors, data exfiltration. That shouldn't put you off agentic AI. It should change how you set it up. The failure modes have names now: Skill poisoning. A skill is not a template, it's a set of instructions your agent obeys. A poisoned skill hides extra instructions inside a useful one. One researcher found a marketplace skill that worked perfectly as a calendar integration while copying data to an external server in the background. It passed testing because it did what it claimed. It just also did something it didn't claim. Tool drift. It was safe when you approved it. Then it updated. Most marketplaces scan skills at creation, not on every update, and on at least one major platform a same-name skill silently replaces the one your team already trusts. Yesterday's review does not cover today's version. And the safety nets are thinner than they look. Install counts can be spoofed. Researchers bypassed the major skill scanners in under an hour. The description is marketing; the SKILL.md is the truth. None of this is a reason to sit out. The organisations getting real value from agents treated governance as part of the build: Read-only by default. Humans approve anything state-changing. Access matched to the job. A writing skill never needs your inbox. Every skill reviewed like untrusted code, and re-reviewed the moment it updates. Outbound traffic watched. Audit trails that name the agent, the policy, and the human who authorised it. The gap between an AI incident and an AI advantage is the governance layer underneath. Set your business up for success, not security risk. That's the work we do at Possibl AI. Build your brain with us - www.possibl.ai #AgenticAI #AIGovernance #AISecurity

"The AI powered brain is the most significant single change in Fundtap, ever." Matt Peacey, Founder and Director of FundTap, on the AI Brain Possibl built for his team this year. FundTap is a 10-person invoice finance business in Nelson, funding small businesses across New Zealand and Australia. Eight of those ten people now have an active project running through the Brain, 26 active projects in all, and 58% of them coordinate two or more people off a shared plan instead of a meeting. Kicked off in May. Handed over in June with the operations and run manuals, and FundTap's own Steward has run it since. Give a language model your company documents and it will answer confidently, whether those documents are current or not. Let everyone edit them and your single source of truth stops being single. Let an agent write to them unsupervised and it will quietly rewrite your strategy to match last week's sales calls. So we put the governance in the architecture, where it can actually be enforced. → Facts and strategy live in separate layers → One Steward writes. Humans and AI propose → 15 proposals decided a week, 98 since day one → Live data stays in HubSpot, Xero and ClickUp → Connectors are read-only by default → Every request uses FundTap's own identity Instructions are guidance, architecture is control. An agent can ignore an instruction file. It cannot ignore a permission. The full case study has the architecture, the four enforcement layers in order of strength, how the connectors and the Azure environment are built, and what FundTap's Steward actually does each week. → https://lnkd.in/eVKi38h7 Thank you to Matthew Peacey and Helena Howley at FundTap for backing this, and for letting us show the real numbers. #AIGovernance #EnterpriseAI #AIinNZ #PossiblAI

🌿 Aotearoa AI Hackathon Festival 2026 Mentor Spotlight 🌿 Building impactful, human-centered tech doesn’t happen in a vacuum—it takes guidance from standard-setting industry leaders! We are thrilled to introduce our first cohort of incredible mentors guiding participants at this year’s Aotearoa AI Hackathon Festival: ✨ Yesha Kaniyawala – AI Software Engineer ✨ Prasanth Pavithran (MBA) – Senior Business Analyst ✨ Mark Modricker – Senior ICT Manager From technical engineering and architecture to strategic business analysis and IT operations, our mentors bring a wealth of practical experience to help participants turn bold ideas into viable, real-world AI solutions. 🚀 📅 Event Details: Dates: August 7, 2026 (5:00 PM – 8:00 PM) & August 8, 2026 (7:30 AM – 8:00 PM) 📍Location: AUT WG Building, City Campus, Auckland Proudly brought to you by SheSharp in partnership with Auckland University of Technology (AUT).