OpenAI Launches Daybreak as Google Confirms First AI-Written Zero-Day: Welcome to the AI Cybersecurity Arms Race

Two stories broke this week, and they’re telling us the same thing: the AI cybersecurity arms race is no longer theoretical. It’s happening, right now, on both sides of the line.

OpenAI launched Daybreak, a dedicated AI cybersecurity platform combining GPT-5.5-Cyber with Codex Security. Meanwhile, Google Threat Intelligence Group confirmed it stopped the first-ever AI-developed zero-day exploit — a Python script with a hallucinated CVSS score and textbook LLM formatting, aimed at mass-exploiting a 2FA bypass. These aren’t unrelated events. They’re opposite ends of the same fuse.

🔍 THE BOTTOM LINE

The AI cybersecurity arms race just went hot. OpenAI and Anthropic now have competing security AI products, and the first AI-written zero-day exploit has been confirmed in the wild. The question is no longer “if” AI will change cybersecurity — it’s whether defence can keep pace with offence.

What Is OpenAI Daybreak?

OpenAI’s Daybreak is an AI security initiative that detects and patches vulnerabilities before attackers find them. It uses Codex Security (launched March 2026) to build a threat model from an organisation’s codebase, identify attack paths, validate likely vulnerabilities, and automate detection of the highest-risk ones.

What makes it different? Daybreak isn’t a single model. It combines GPT-5.5-Cyber, GPT-5.5 with Trusted Access for Cyber, Codex, and security partner integrations. OpenAI says it’s working with “industry and government partners” while preparing to “deploy increasingly more cyber-capable models.”

This directly counters Anthropic’s Claude Mythos and Project Glasswing — which were announced as “too dangerous to publicly release” before being breached by unauthorised parties on launch day. OpenAI is explicitly positioning Daybreak as the safer, more responsible alternative.

The irony? Just last month, GPT-5.5 matched Mythos on cybersecurity benchmarks, and Sam Altman called Mythos’ fear-based positioning exactly what it was: marketing. Now OpenAI has its own security product, and the gloves are off.

The First AI-Written Zero-Day: What Google Found

Google Threat Intelligence Group’s report is a milestone. Here’s what they found:

The target: An open-source web-based system administration tool with a hardcoded 2FA trust assumption
The exploit: A Python script exploiting a high-level semantic logic flaw — the kind an AI is particularly good at identifying
The AI fingerprints: A hallucinated CVSS score (something LLMs do when they fabricate metadata), structured “textbook” formatting consistent with training data, and an exploit methodology that looked AI-refined
The intent: Mass exploitation — the attackers planned a widespread campaign using this zero-day
The outcome: Google disrupted the exploit, but notes this is a “first confirmed” — not a “last”

Importantly, Google says it does not believe its own Gemini was used. The AI in question was likely an open-source or third-party model, jailbroken through “persona-driven” prompting — instructing the AI to roleplay as a security expert to bypass its safety guardrails.

Attackers are also feeding entire vulnerability repositories into AI models and using automation tools to “refine AI-generated payloads within controlled settings to increase exploit reliability prior to deployment.” They’re using AI to make better exploits, faster.

The Escalation Nobody Wanted

Read these two stories side by side and a clear picture emerges:

Defence is getting serious. OpenAI Daybreak, Anthropic Project Glasswing, Google’s threat intelligence — three of the biggest AI companies now have dedicated security AI products. GPT-5.5-Cyber and Mythos are in a direct capability race, with each new benchmark showing them roughly neck-and-neck.

Offence is already here. The first AI-written zero-day has been confirmed. Not a proof of concept. Not a researcher demo. A real attack, aimed at real systems, disrupted before it could hit. And Google itself says attackers are getting better at using AI to find vulnerabilities.

The timeline is compressing. Small AI models matched Mythos on cybersecurity just a month ago — showing that security-capable AI isn’t limited to the labs with billion-dollar budgets. Open-weight models with security training are within reach of anyone. The democratisation goes both ways.

Remember when Mythos triggered an emergency Treasury meeting because cyber stocks collapsed on the news? That was five weeks ago. The industry has moved at warp speed since.

What This Means for NZ

New Zealand’s cybersecurity landscape is relevant here for a few reasons:

SMB exposure. NZ’s economy is dominated by small and medium businesses — exactly the kind that are least prepared for a new generation of AI-powered attacks. A mass-exploitation zero-day targeting open-source admin tools would hit our SMB sector hard.
Government posture. The AI Blueprint for Aotearoa was refreshed by AI Forum NZ just last week, but it’s light on cybersecurity specifics compared to what’s now unfolding. The UK just launched a national AI chatbot, the US is grappling with AI-powered cyber threats, and Australia is advancing its own AI safety framework. NZ needs to be thinking about this now — not after the first major incident.
Critical infrastructure. Our power grid, healthcare, and banking systems rely on the same open-source tools and supply chains that AI-assisted attacks are targeting. We’re not insulated by distance.

❓ Frequently Asked Questions

Q: Is AI cybersecurity a real product category now, or just marketing? It’s real. OpenAI Daybreak and Anthropic Project Glasswing are both generating revenue and are taken seriously enough that Google is publishing detailed threat intelligence specifically about AI-powered attacks. The category exists because the threat exists.

Q: Does this mean AI is now a net negative for security? No — but the balance matters. AI-powered defence is genuinely improving: automated patching, faster vulnerability detection, better threat modelling. The problem is that AI-powered offence is accelerating faster than defence can adapt. The net effect depends on whether institutions adopt defensive AI quickly enough.

Q: What can I do as an individual? Keep your software updated — this zero-day targeted a 2FA bypass, and patches are the first line of defence. Use strong, unique passwords. Enable multi-factor authentication where available. And don’t assume “I’m not important enough to target” — mass-exploitation campaigns don’t discriminate.

Q: What should NZ businesses do? Audit your supply chain. Know what open-source tools your infrastructure depends on. If you’re a small business, your managed IT provider should have a vulnerability management process. If you don’t have one, that’s your starting point.

🔍 THE BOTTOM LINE

One week. Two stories. Same takeaway: the AI cybersecurity arms race has gone hot on both sides. OpenAI launched the shield; attackers deployed the sword. The first AI-written zero-day has been confirmed in the wild, and defence is scrambling to keep up. This isn’t a future problem — it’s today’s news, and tomorrow’s is likely worse.

The race between AI-powered security and AI-powered attacks is the most important technology story of 2026. And it’s barely started.