AI and cloud briefing: OpenAI foils a 15,000-account bid to steal AI reasoning, Google launches Gemini 4 Argon, and Broadcom backs Anthropic with $42 billion

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: OpenAI shuts down a 15,000-account campaign trying to steal its models’ hidden reasoning

OpenAI has disclosed that it disrupted a large, coordinated campaign aimed at extracting the hidden “reasoning” its models use before producing a final answer. The activity started slowly on 1 July 2026, then spiked sharply on 24 and 25 July with around 16,000 requests from more than 4,000 accounts, before investigators traced a wider network of over 15,000 accounts in total. OpenAI says it attributed a core part of the activity to individuals associated with Moonshot AI, the developer of the Kimi AI model, though it could not confirm that every account belonged to a single actor. The attackers reportedly used coordinated prompts and cross-conversation tricks, including copying encrypted reasoning output from one conversation and feeding it into a separate conversation with instructions to decrypt it. OpenAI says it fully disrupted the operation by 28 July, then banned or restricted the accounts involved, tightened sign-up and infrastructure controls, closed off the replay routes used to recover encrypted reasoning, added new safeguards to catch hidden reasoning leaking through streamed output, and shared details with other AI developers through the Frontier Model Forum. More detail is available from Cyber Security News’ report.

Why this matters: AI providers now treat a model’s internal reasoning as valuable intellectual property worth defending at scale, not just its answers. For UK businesses building products on top of these models, it is a reminder that usage terms around reasoning output exist for a reason, and that providers are actively watching for and shutting down attempts to copy a competitor’s model this way.

Artificial intelligence: Google gives cyber defenders first access to Gemini 4 Argon

Google has launched Gemini 4 Argon, which it describes as its strongest model yet for complex, long-running professional work in software engineering, enterprise knowledge tasks and cyber defence. It raises the output limit to 1 million tokens, up from 64,000 on its predecessor, and posts strong scores on independent benchmarks: 77.9% on DeepSWE v1.1 for real-world software engineering tasks, first place on AutomationBench at 51.3%, a lead on the Vals Index of economic impact across finance, coding, legal and tax work, 91.7% on LVBench for long video understanding, and a tied first place on CWE-bench v1 at 68% for vulnerability remediation. Rather than a broad public launch, Google is giving trusted cyber defenders first access through its Fairwind Program, and is working through the US government’s voluntary pre-release access process before opening the model up to developers, businesses and consumers more widely. Launch pricing is $2 per million input tokens and $10 per million output tokens. Google says “Argon is fundamentally changing the way we work and build” internally. More detail is available from Google’s own announcement and SecurityWeek’s coverage.

Why this matters: Google restricting its most capable model to vetted security teams first, rather than releasing it broadly on day one, shows how seriously frontier labs now take the risk that a powerful coding and reasoning model could be misused. It is also a preview of what is coming to everyone else, so UK security and engineering teams should watch how this staged rollout plays out over the coming months.

Cloud infrastructure: Broadcom agrees to lend Anthropic up to $42 billion for AI chip capacity

Broadcom has agreed to provide Anthropic with financing of up to $42 billion, structured as convertible debt that Anthropic could later turn into equity, with Broadcom able to designate a financing partner for the arrangement. The money is earmarked to help fund Anthropic’s lease of tensor processing unit (TPU) computing capacity from Google and Broadcom, and covers roughly a third of Anthropic’s total five-year TPU commitment of $125.2 billion. Broadcom is expected to become Anthropic’s largest compute supplier in 2027, and separately projects around $115 billion in AI semiconductor revenue for its 2027 financial year, rising to $230 billion in 2028. Anthropic has flagged a potential conflict of interest in the deal, since Broadcom is acting as both hardware supplier and financier, which could influence Anthropic’s future computing access and pricing. The arrangement echoes a financing approach Nvidia has used with its own customers. More detail is available from Pulse2’s coverage.

Why this matters: this is one of the largest financing deals yet in the AI infrastructure build-out, and it shows chip suppliers increasingly bankrolling the AI labs that buy their hardware, tying supplier and customer together more closely than a standard purchase order. For UK businesses, deals like this underpin the compute capacity that eventually shows up in AI API pricing and cloud costs, so it is worth keeping an eye on whether arrangements like this keep compute costs stable or add new risk to the supply chain.

The takeaway

Today’s stories share a common thread: AI providers are working hard to protect, control access to, and finance what has become their most valuable asset, the model itself. OpenAI moved to stop its reasoning being copied at scale, Google is restricting its newest model to vetted defenders before a wider release, and Broadcom is willing to put up tens of billions of dollars to keep Anthropic supplied with compute. For UK businesses, it is a reminder that the economics and security of frontier AI are both still being worked out in real time. 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.