What's New in AI: Frontier Models Go Public — 13 July 2026

The frontier models came off the leash this week. OpenAI's most capable system finally reached the public after a government safety gate lifted, xAI dropped a rival coding model days later, and the world's first fully autonomous AI ransomware operation showed up in the wild — all while the United Nations opened its first-ever global forum on how to govern any of it. Here's your plain-English roundup of what happened and why it matters for your business.

GPT-5.6 goes fully public — and Grok 4.5 crashes the party

The AI you can actually use took a big step up this week. On 9 July, OpenAI opened its GPT-5.6 family (the Luna, Terra and Sol tiers) to the general public and made it the default in ChatGPT, ending a government-imposed limited preview that had kept the model behind a White House cybersecurity review for nearly two weeks. OpenAI was pointed about it, saying that kind of pre-release government access “shouldn't become the long-term default”. A day earlier, xAI launched Grok 4.5, a model trained on trillions of Cursor coding tokens, priced aggressively at roughly US$2 per million input tokens and shipped straight into Cursor and the xAI console.

Why it matters: the practical takeaway isn't the leaderboard — it's that top-tier reasoning and coding are now cheap, fast and available to everyone at once. When two of the strongest models on the market land in the same week and compete on price, the cost of building genuinely useful AI features keeps falling. That's exactly the moment small and mid-sized businesses can move: the barrier is no longer access to the model, it's knowing how to wire it into your workflow. That's the work our AI development team spends its days on.

The first fully autonomous AI ransomware attack lands

Security researchers at Sysdig disclosed JADEPUFFER, which they assess to be the first documented ransomware operation run end-to-end by a large language model rather than a human. The agent broke into an internet-facing server through a known Langflow vulnerability, hunted for credentials, moved laterally to a production database, planted a backdoor, and then encrypted and destroyed data — narrating its own reasoning in the code the whole way. In one telling moment, it went from a failed login to a working multi-step fix in 31 seconds, far faster than a human operator could diagnose and repair the error.

Why it matters: none of the individual techniques were novel — the attack leaned on years-old, unpatched flaws. What's new is that an AI agent chained them together without a skilled human at the keyboard, dropping the skill floor for running an attack to roughly the cost of renting a model. Last week we covered how cheap AI agents went mainstream; JADEPUFFER is the mirror image of that story. The defensive basics — patching, scoping credentials, locking down internet-facing databases — just became urgent rather than optional.

The UN opens its first global AI governance forum

Governments, tech firms, academics and civil society gathered in Geneva on 6–7 July for the first UN Global Dialogue on AI Governance, the earliest attempt to put all nations around one table on the technology. It followed a preliminary report from the UN's new 40-member Independent International Scientific Panel on AI, co-chaired by Yoshua Bengio and journalist Maria Ressa. Bengio's warning was blunt: science “cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm”, whether on its own or in the hands of malicious users. Secretary-General Guterres pushed four priorities, including common safety standards and human-rights red lines.

Why it matters: a “dialogue” produces a co-chairs' summary, not binding law, so nothing changed overnight. But it signals where the regulatory wind is blowing — toward shared baseline standards that will eventually shape what businesses can and can't deploy. For anyone building on AI now, the message is to design for transparency and human oversight from the start, because that's the direction the rules are heading.

Australia gets more hands-on with AI

Closer to home, the Federal Government shifted toward a more interventionist stance, tilting its policy to weigh community benefits more heavily against the interests of data-centre operators. Under rules being considered, new data centres would need to show clear evidence of local jobs, investment and spending before getting the green light. Assistant Minister Andrew Charlton warned that powerful models are already doing things their creators never intended, and the government has stood up an AI Safety Institute under the National AI Plan as a national model-testing capability. Australia, Canada and India also signed a technology and innovation partnership (ACITI) during the month.

Why it matters: Australia is edging away from a purely light-touch approach without going full-EU. For local operators, that means AI safety and community accountability are becoming part of the compliance conversation, not just an ethics footnote. If you're adopting AI tools, expect customers, partners and regulators to increasingly ask how you're using them — and to reward businesses that can answer clearly.

Humanoid robots march toward the public markets

Physical AI had a landmark week as three humanoid-robot stories converged. Agility Robotics agreed to go public via a US$2.5 billion SPAC, becoming the first US-listed pure-play humanoid company, backed by more than US$300 million in booked orders. In China, Unitree cleared final registration for a roughly US$618 million Shanghai STAR Market IPO after a record review, and Tesla began converting a line toward Optimus production, targeting hundreds of units a week through the year — though those 2026 robots are for internal factory use, not sale.

Why it matters: humanoids are moving from lab demos to balance sheets, and public listings force the discipline of real revenue and real deadlines. We're still years from robots being commonplace in most workplaces, but the capital is flooding in and the manufacturing lines are being built. It's a useful reminder that “AI” increasingly means hardware in the physical world, not just chatbots on a screen.

The funding taps stay open — for applied AI

Investors kept writing large cheques, and the money is getting more practical. Infrastructure firm Together AI reportedly closed an US$800 million round for its platform that lets enterprises train and run open-source models, while agentic decision platform Taktile raised US$110 million to automate regulated workflows such as loan approvals and claims triage. Smaller rounds flowed to hardware and robotics plays too, per the week's funding roundup. Notably, nearly 88% of AI startup funding in 2026 has gone to US-based companies.

Why it matters: the froth is shifting from “build a bigger model” to “put AI to work inside real business processes”. That's good news for everyone downstream: more capital chasing applied AI means better, cheaper tools for automating the unglamorous workflows that actually eat your week. The concentration of funding in the US, though, is a quiet warning for the rest of the world about who owns the infrastructure.

Quick hits

• Google speeds up creative AI: NanoBanana 2 Lite now generates images in under four seconds from about US$0.034 per 1,000, and Veo 3.1 stands as a leading video model. See the month's release tracker.
• Meta ships its first paid model: Muse Spark 1.1 launched on 9 July as a commercial offering, a notable pivot for a company known for open weights (details).
• Chinese open models surge in US enterprises: Chinese models have reportedly accounted for 30–46% of US enterprise API token usage since February, with Z.ai's GLM 5.2 growing fast (report).
• OpenAI adds voice and work tools: alongside GPT-5.6, OpenAI introduced “GPT-Live” conversational models that listen and speak at once, plus a ChatGPT Work tool (Axios).
• Smart glasses raise big: Even Realities banked US$150 million for camera-free smart glasses, a sign wearables remain a hot AI form factor.

Looking ahead

Watch for Grok 4.5 to reach the EU in mid-July, and for the first independent benchmarks pitting GPT-5.6 Sol against Grok 4.5 on real coding work rather than vendor slides. Expect the JADEPUFFER disclosure to accelerate “agentic threat” guidance from security vendors, and keep an eye on whether the UN's Geneva summit produces any concrete follow-through. If you want a head start, our library of AI tools and resources is a good place to see what's worth trying now.

That's the week in AI — the models got cheaper and more capable, the risks got more concrete, and the rule-makers finally sat down together. Check back next week for the next instalment of what's new in AI, and what it means for your business.