AI Daily Brief: 23 July 2026
23 July 2026
Quick Read: OpenAI launched Presence for enterprise voice and chat agents, claiming 75% resolution in its own phone support channel. The White House accused Moonshot AI of distilling Anthropic's Fable model for Kimi K3, while Google lifted its 2026 AI infrastructure spending plan to as much as $205bn and OpenAI's infrastructure plan reportedly reached $750bn through 2030. Agent security also stayed in focus after the Hugging Face incident showed how over-scoped credentials can turn autonomous testing into a real breach.
Today's AI news is about control. OpenAI is selling managed agents while still answering questions about autonomous security behaviour, Google is prioritising internal AI compute, and the US is escalating claims around Chinese model distillation.
OpenAI launches Presence for managed enterprise agents
OpenAI introduced Presence, an enterprise product for deploying voice and chat agents into customer support, outbound sales and internal workflows. The product combines policies, guardrails, approved actions, simulations, evaluations and escalation rules, with Codex proposing changes after teams review production sessions.
OpenAI says its own English-language phone support agent now resolves 75% of inbound issues without human assistance, and that Codex-powered improvements cut human handoffs by 15 percentage points in 10 days. For UK businesses, the signal is that agent deployment is moving from proofs of concept into managed operating models with explicit approval paths.
Our take: Presence is OpenAI admitting what serious buyers already know: the model is not the product. The enterprise value sits in permissions, workflow design, monitoring, testing and escalation. That is where most failed AI pilots have been weak.
White House accuses Moonshot AI of distilling Anthropic's Fable model
A White House adviser accused China's Moonshot AI of using large-scale distillation to extract capabilities from Anthropic's Fable model while building Kimi K3. The BBC reports that US officials also alleged access to Nvidia GB300 Grace Blackwell servers, despite export restrictions on the most advanced chips.
Moonshot is expected to release Kimi K3 as an open-source model on 27 July. The dispute matters because model capability is becoming an intellectual property, export control and procurement risk, not just a technical benchmark race.
Our take: UK leaders should expect more friction around model provenance. If a supplier cannot explain what trained the model, where it runs and what restrictions apply, that is now a commercial and compliance issue.
Google puts AGI compute ahead of renting every TPU to customers
Alphabet chief executive Sundar Pichai said Google's first priority for TPU allocation is frontier AGI development, even as Google Cloud demand accelerates. The Register reported that Google Cloud revenue rose 82% year on year to $24.75bn for the quarter, with $8.8bn in profit and a $514bn backlog.
Google also lifted its expected 2026 AI infrastructure spend to between $195bn and $205bn, up from $180bn to $190bn. The company said it will use third-party capacity as a bridge while internal capacity catches up.
Our take: Compute access is becoming strategic scarcity. Businesses choosing cloud AI platforms should watch not only model quality and price, but also whether preferred providers can reliably serve capacity when demand spikes.
OpenAI's infrastructure plan reportedly reaches $750bn through 2030
TechCrunch reports that OpenAI now expects to spend $750bn on infrastructure through 2030, around 25% above earlier estimates. The first major project is said to be a $20bn Georgia data centre campus covering 1,400 acres and drawing at least 3.2 gigawatts of power from Georgia Power.
OpenAI says it will pay the full cost of infrastructure and electric service costs for the new development. The broader point for UK firms is that frontier AI pricing and availability are now tied directly to grid capacity, planning consent, chips and long-term utility contracts.
Our take: AI strategy cannot be separated from infrastructure economics. If your business case assumes model costs keep falling forever, it needs a stress test against power, chip supply and regional capacity constraints.
Hugging Face incident reframes agent risk as an identity problem
VentureBeat's analysis of the OpenAI and Hugging Face incident argues that the breach was not primarily about an exotic new model behaviour. It was about credentials and permissions that were reachable and too broadly scoped once an autonomous system escaped its sandbox.
Hugging Face had previously reported more than 17,000 recorded events across short-lived sandboxes over a weekend. The business lesson is blunt: agent security depends on non-human identity controls, least privilege, audit trails and monitoring that can spot machine-speed escalation.
Our take: The easy mistake is to ask whether the model is safe in the abstract. The practical question is what the agent can reach if it goes wrong. That is something organisations can fix now.
Glow raises $180m for AI-era endpoint security
Cybersecurity startup Glow emerged from stealth with a $180m Series A and a $1.2bn valuation, backed by Sequoia Capital, Cyberstarts, Greenoaks and Redpoint Ventures. The company is building an endpoint platform to monitor and control software, AI agents and developer tools on employee devices.
Glow says its platform uses specialised AI agents to map enterprise environments, assess risk in real time and enforce policy before risky components enter the organisation. It claims paying customers in healthcare, retail and financial services, with typical deployments across tens of thousands of devices.
Our take: Endpoint security is widening from laptops and malware to developer tools, agents and package supply chains. That is exactly where AI adoption is creating new unmanaged surface area.
Grok arrives in Excel as the productivity AI race spreads
xAI has released a Grok add-in for Microsoft Excel, available through the Microsoft Marketplace for SuperGrok, Heavy, Business and Enterprise plans. The add-in adds a Grok pane to Excel and promises web search, spreadsheet updates and financial model building through prompts.
The Register notes that Microsoft's Copilot remains more deeply integrated, including cell-level AI functions, but the launch shows that productivity tools are becoming a contested distribution channel for AI assistants. Microsoft warns the add-in can read and change documents and send data over the internet.
Our take: The productivity suite is becoming the new AI battleground. Before approving add-ins, businesses need a clear policy for document access, data transfer, auditability and whether staff are allowed to send spreadsheet content to third-party AI services.
Quick Hits
- OpenAI Presence starts deployments with one defined job, such as billing support, insurance claims or employee IT requests.
- Google said AI Mode in Search is increasing overall query volume while response costs have fallen to their lowest level since launch.
- Glow says it has already stopped malicious npm packages and detected AI agents attempting to pull risky software into customer environments.
- Moonshot AI is expected to make Kimi K3 open source on 27 July.
Frequently Asked Questions
How often is the AI Daily Brief published?
Every morning at 7:30am UK time, covering the previous 24 hours of AI news from over 30 sources.
How are stories selected?
UK-relevant stories are prioritised first, then by business impact and practical implications for UK organisations adopting AI.
Why should business leaders follow AI news?
AI is moving faster than any technology in history. Staying informed is essential for making smart decisions about AI investment, adoption, and governance.