· artificial intelligence · 6 min read

AI and cloud briefing: Anthropic's Claude tops the AI rankings as OpenAI locks down a 'Critical'-rated cyber model, AMD's record quarter meets new US chip curbs on China, and a Zyxel zero-day hits nearly 1,000 switches

A fast-moving week in AI saw Anthropic's Claude Fable 5.2 top an independent intelligence ranking while OpenAI restricted its new GPT-6 Cyber model after rating it 'Critical' risk. Separately, AMD posted a record quarter even as the US tightened chip export curbs on China. And Zyxel has patched a zero-day that let attackers compromise close to 1,000 network switches worldwide.

A fast-moving week in AI saw Anthropic's Claude Fable 5.2 top an independent intelligence ranking while OpenAI restricted its new GPT-6 Cyber model after rating it 'Critical' risk. Separately, AMD posted a record quarter even as the US tightened chip export curbs on China. And Zyxel has patched a zero-day that let attackers compromise close to 1,000 network switches worldwide.

Here is our latest round-up of what’s changing in AI and cloud for UK businesses, with links to the original sources so you can read further.

Lead story: the AI model race hits a new pace, and safety limits are starting to bite

The last two weeks have brought a burst of new AI model releases from almost every major lab. Anthropic released Claude Fable 5.2 on 9 November, and it has since topped the independent Artificial Analysis Intelligence Index with a score of 65, ahead of Google’s Gemini 4 Carbon (60) and Anthropic’s own Claude Opus 5.5 (59). Alongside it, Anthropic launched Claude Mythos 5.2, a more capable variant restricted to vetted cybersecurity and life sciences organisations. Google followed on 12 November with Gemini 4 Hydrogen, a model built for everyday coding and agentic work, offered through its Antigravity coding tool at four selectable “thinking” levels. Meta, Baidu, Thinking Machines Lab and Google DeepMind also shipped updates in the same window, including DeepMind’s Gemini Robotics 2.5, which reportedly achieved a 96% success rate sorting parts in a pilot on Mercedes-Benz production lines.

The most striking move came from OpenAI. On 5 November, it released GPT-6 Cyber, a model trained for offensive and defensive cybersecurity work, but gave access only to approved customers in its restricted Daybreak Red programme. In internal testing, the model found 81 previously unknown high-severity vulnerabilities in Firefox and built 948 working exploits against the Linux kernel. OpenAI rated it “Critical” under its own Preparedness Framework, the first time any of its models has reached that level, which is why public access was withheld. More detail is available from PYMNTS’s coverage of the rankings and OpenAI’s restricted model and OrcaRouter’s breakdown of Gemini 4 Hydrogen.

Claude Fable 5.2 tops the AI rankings as OpenAI locks down its Critical-rated cyber model

Why this matters: a model capable of finding dozens of real, unknown software flaws is a genuine step forward for defenders, but OpenAI’s own “Critical” rating shows that the same capability could do serious harm in the wrong hands, which is exactly why it has not been released publicly. For most UK businesses, the practical point is simpler: the gap between what the newest AI models can do and what your team or suppliers are actually using is widening quickly, so it is worth checking which model versions your tools rely on and whether vendors are being clear about what capabilities, and what restrictions, come with them.

Cloud and chips: AMD’s record quarter meets tighter US curbs on China

AMD reported record third-quarter results on 3 November, with revenue up 45% to $13.4 billion and data-centre sales nearly doubling to $8.5 billion, driven by AI chip demand. It forecast around $17 billion in fourth-quarter revenue and, on 5 November, agreed to pay about $2.9 billion for Toronto-based AI chip start-up Taalas. The good news was tempered by wider turbulence in the chip supply chain. Around the same time, roughly 600,000 Nvidia AI chips were reportedly held up in Hong Kong awaiting Chinese import approval, while the US ordered Taiwan’s TSMC to halt advanced chip exports to China and withheld export approvals for SK hynix’s Chinese memory plant. Separately, Marvell’s proposed $115 billion all-stock acquisition of Lumentum has seen both companies lose a combined $91 billion in market value on investor doubts, and Intel’s shares fell 28% in a single day despite 21% revenue growth, after its guidance disappointed the market. More detail is available from Distill Intelligence’s semiconductor and AI chips briefing.

AMD's record quarter came alongside new US curbs on chip exports to China

Why this matters: the chip market underpins almost every AI product a business might buy, from cloud compute to the hardware inside new laptops, and this week shows how unsettled that supply chain still is. Export restrictions, stalled shipments and sharp swings in investor confidence can all affect pricing and availability further down the chain, so it is worth building some flexibility into any technology plans that assume AI compute will keep getting cheaper and more available at the same pace it has recently.

Cyber security: a Zyxel zero-day compromises nearly 1,000 network switches

Zyxel has patched a zero-day vulnerability, tracked as CVE-2026-107441 with a severity score of 8.8, in its GS1900 series network switches. The flaw, an out-of-bounds write, let an unauthenticated attacker on the local network run operating system commands through a specially crafted request. Before the patch, it was used to compromise close to 1,000 devices worldwide, stealing configuration files and credentials. The activity has been linked to a Chinese-speaking group tracked as Kapibala, also known as Red Heron, which deployed a backdoor called JITTERLY and a rootkit called SIXZUT. In the same week, researchers also flagged active exploitation attempts against a separate OpenSSH authentication bypass, nicknamed “SSHole” and tracked as CVE-2026-123158, which can allow sign-in to accounts with no password set on certain Linux configurations. More detail is available from The Hacker News’s weekly recap covering both flaws.

A Zyxel zero-day compromised nearly 1,000 network switches worldwide

Why this matters: network switches and SSH access sit at the heart of most office and data-centre networks, and both of this week’s flaws allowed attackers to get in before anyone had a chance to patch. If your business, or an IT provider acting on your behalf, runs Zyxel GS1900 switches or Linux systems with OpenSSH, this is worth checking promptly: confirm firmware and package updates are applied, and review device and access logs for anything unusual, since attackers who gain a foothold this way often stay quiet while they look for a bigger target.

The takeaway

This week’s stories point to the same idea from three different angles: AI capability, computing hardware and basic network security are all moving fast, and all three carry real limits that are becoming more visible. OpenAI’s own “Critical” safety rating for GPT-6 Cyber is a reminder that more capable AI brings more responsibility, not less. The chip market’s mixed results show that AI’s physical supply chain is still far from settled. And the Zyxel and OpenSSH flaws show that unpatched, everyday infrastructure remains one of the most common ways attackers get in. We can help you work through what applies to your business. Get in touch.


AI Agent based research and content, AI can do mistake, please help us to improve.

Back to Blog