AI Wrote a Zero-Day Exploit — Why AI Literacy Just Got Urgent
Google confirmed the first zero-day exploit developed with AI assistance. Attackers used AI to write code that would bypass two-factor authentication. Google caught it before deployment but warns this is the new normal. The exploit bore hallmarks of AI-generated code: a “hallucinated CVSS score” and overly neat formatting.
Why it matters for education: If AI can now write weaponised code, AI literacy isn’t just about understanding chatbots — it’s about understanding what AI can do to infrastructure. Students entering tech, finance, healthcare, or government need to understand AI-enabled threats as a core competency. The “AI 101” curriculum just got harder, and a lot more important.
KAIST Breakthrough: Teaching AI Models to Say “I Don’t Know”
South Korean researchers at KAIST developed a method where neural networks undergo “warm-up” training with random noise inputs, helping them recognise when they’re dealing with unfamiliar knowledge. The approach directly addresses the fundamental problem of AI overconfidence and hallucination — models that confidently answer questions they shouldn’t.
Why it matters for education: This is one of the most important alignment developments of 2026, and it has direct implications for education. If AI tutors can recognise when they don’t know something, they stop confidently teaching wrong answers. The current generation of AI tutoring tools are confidence machines — they answer everything. Teaching models uncertainty is the single most important capability for AI in education.
Youth AI Safety Institute Launches With Major Backing
EU Commission President Ursula von der Leyen, Hillary Clinton, and Danish PM Mette Frederiksen have backed a new Youth AI Safety Institute focused on protecting children from AI harms. The institute aims to develop standards and frameworks for “childproofing” AI systems.
Why it matters for education: The “youth” focus is significant. While most AI safety work targets existential risk or corporate misuse, this institute targets the everyday harms that affect young people — algorithmic manipulation, deepfake targeting, and AI-assisted grooming. Expect this to influence AI-in-schools policy globally, including NZ.
China’s AI Agent Rules Include Education-Specific Requirements
China’s draft AI agent regulations explicitly mention education tasks like “marking homework” and “evaluating employee performance” as domains requiring human oversight. The rules require that users retain the right to know and the final decision-making power over autonomous agent decisions.
Why it matters for education: China is the first jurisdiction to specifically regulate AI agents in education. The “human in the loop” requirement for marking homework is notable — it suggests Beijing isn’t comfortable letting AI autonomously evaluate students. NZ schools experimenting with AI marking should take note: the regulatory direction globally is toward requiring human oversight of AI in assessment.
OpenAI Gives Vetted Researchers Access to GPT-5.5-Cyber
Following the Mythos cybersecurity earthquake, OpenAI is giving vetted security researchers access to GPT-5.5-Cyber for cybersecurity work. The model is optimised for defensive security research, not offensive capability.
Why it matters for education: The AI cybersecurity race has an education dimension. Universities with cybersecurity programmes need access to these tools to train the next generation of security professionals. If only vetted researchers can access frontier cyber AI models, academic programmes will need new partnerships and access arrangements. NZ’s cybersecurity education is already thin — this makes it thinner unless universities move fast.
🔍 THE BOTTOM LINE
The big theme this week: AI’s capabilities are outpacing our ability to teach about them. AI writes exploits, but most people can’t read code. AI models hallucinate confidently, but we’re only now learning to teach them uncertainty. AI agents are marking homework in China, but we don’t have frameworks for when that’s acceptable. The education system’s biggest AI challenge isn’t “how do we use AI?” — it’s “how do we understand AI well enough to live with it?”