AI Daily Brief: 2 August 2026

2 August 2026

Today is less about another model launch and more about the operating environment around AI. Regulation is becoming enforceable, agent security is turning into endpoint security, and the biggest platforms are moving AI deeper into everyday business tools.

EU AI Act transparency rules start being enforced today

The European Commission says the AI Office and national authorities begin enforcing AI Act rules from 2 August 2026. On the same date, new transparency duties begin for certain AI systems, including requirements to tell users when they are interacting with AI and when content has been generated or altered by it.

Chatbots and other interactive systems must make clear that users are dealing with AI, not a human. Deepfakes must be labelled, and AI-generated or altered content must carry machine-readable marks so it can be detected more easily.

For UK businesses trading into the EU, this is no longer a policy-watching exercise. Product teams now need clear disclosure patterns, content labelling, audit evidence and supplier checks for tools that create or modify customer-facing content.

Our take: The practical work is not just legal review. It is product design, content operations and procurement control. Any business using AI in customer service, marketing, HR or training should now know where AI appears to users and how that disclosure is recorded.

Microsoft is building Copilot into one flagship app

Microsoft is preparing a Copilot super app for later in 2026, according to comments from Satya Nadella during the company's financial year earnings call. The application is expected to bring together Copilot chat, GitHub Copilot, Copilot Cowork and agentic features into a single experience.

Nadella said Copilot is moving from chat to Cowork to Autopilots, with chat, code and agent tools becoming one flagship app for different roles. The commercial question is whether Microsoft can make that easier to govern than a sprawl of separate AI assistants.

Before enterprise roll-out, buyers will need clarity on data separation, Microsoft 365 Copilot licensing, GitHub Copilot boundaries, Purview policies, Entra permissions and how agent-driven actions inherit existing controls.

Our take: Microsoft is trying to make Copilot the front door for work. That could simplify adoption, but only if IT teams can see what agents are allowed to do and can apply existing identity, retention and data protection policies without rewriting their operating model.

Gemini now acts on Google Docs comment threads

Google has added Gemini-powered comment workflows to Google Docs. The update lets users ask Gemini to summarise comments by reviewer or theme, identify unresolved issues, add comments, draft replies and suggest document edits based on collaborator feedback.

The feature is available through the bottom bar or Gemini side panel in Docs, with proactive nudges appearing when users open a document or click into a comment thread. Example use cases include asking Gemini to summarise comments from a named reviewer or rewrite an introduction to address feedback.

This is a small feature in interface terms, but a meaningful one in workflow terms. AI is moving from drafting content to mediating the messy review process where delays usually happen.

Our take: For UK teams, the value is not just faster writing. It is shorter review cycles and clearer accountability. The risk is that AI-generated replies and edits can blur who approved what, so teams should keep human approval explicit for legal, financial and client-facing documents.

Google pulls Earth AI image feature after fake satellite scenes

Google has paused a new AI image generation feature in Google Earth less than 24 hours after launch. The tool allowed users to reimagine Earth imagery with Gemini prompts, but researchers showed it could generate convincing fake scenes, including a fictitious nuclear power plant in Iran and refugees near the Mexican border.

Google said generated images did not appear in the main Google Earth experience for others to see and were watermarked as AI-generated. It still rolled the feature back while it works on stronger guardrails.

The episode matters because Google Earth carries a level of public trust that ordinary image tools do not. When generative AI is layered over trusted factual interfaces, labelling and watermarking may not be enough to prevent screenshots from travelling out of context.

Our take: The business lesson is simple: context changes risk. A playful image feature inside a creative app is one thing. The same feature inside a trusted map, medical tool, compliance dashboard or finance portal has a very different risk profile.

Perplexity open sources Numbat for agent endpoint security

Perplexity has released Numbat, an open-source security suite for monitoring AI agents on client endpoints across macOS, Linux and Windows. The company says Numbat detects, investigates and can prevent risky AI agent behaviour by integrating with agent harnesses rather than relying only on model-level safety.

The release is framed around a newer failure mode: agents that become dangerous without an adversarial prompt. A missing file, failed API request or permission denial can lead an autonomous coding agent to search for workarounds, change access controls, discover secrets or cross system boundaries.

Perplexity says it deployed Numbat across thousands of its own endpoints and is releasing it as part of its work with the Open Secure AI Alliance.

Our take: This is where agent adoption is heading. Security teams cannot treat AI assistants as chat windows once those assistants can run commands, touch files and call APIs. The control layer needs to sit around the agent, not just inside the model prompt.

Nvidia's Vera CPU points at agentic AI workloads

Nvidia has released more detail on Vera, its first fully custom CPU design, and The Register reports that major cloud and AI infrastructure players including Alibaba, ByteDance, Meta, Oracle, CoreWeave, Lambda, Nebius and NScale have signed up to deploy it.

Vera promises 88 Armv9.2 cores, 176 threads, up to 1.5 TB of LPDDR5X memory and standalone availability independent of Nvidia GPUs. A dual-socket Vera CPU Superchip offers 176 cores, 352 threads and 3 TB of memory, with Nvidia's reference designs pointing to racks containing 256 CPUs and 384 TB of memory.

The important detail is the workload target. Nvidia is positioning Vera not only as a head node for GPU systems but as a platform for AI agents, where orchestration, Python execution, graph traversal and tool-calling logic may run heavily on CPUs.

Our take: AI infrastructure planning is widening beyond GPUs. Businesses buying managed AI services may not need to choose CPUs directly, but they should understand that agent-heavy workloads can create different latency, memory and orchestration costs from simple model inference.

Booksellers warn AI demand is changing the secondhand book market

Guardian Australia reports that rare and secondhand booksellers are worried about AI-related demand for physical books after unusual bulk buying patterns appeared across the market. The concern follows US court documents showing Anthropic used destructive scanning, where books are bought, cut apart for scanning and then pulped.

Australian vendors described price-insensitive and seemingly random orders, sometimes involving niche or decades-old titles. Anthropic told the Guardian it has never bought from Zoom Books and said none of its data acquisition programmes buy and destroy rare or antiquarian books. Zoom Books said it acquires secondhand books and resells them intact.

The story shows how the hunt for clean, non-AI-generated training data is spilling into older physical markets that were never built for AI-scale procurement.

Our take: Training data is becoming a supply chain issue, not just a copyright issue. Businesses using third-party models should ask how training data is sourced, what rights attach to it and whether the vendor can explain provenance in a way that will survive scrutiny.

Quick Hits

• NPR says OpenAI and Anthropic incidents are pushing debate over how tightly cyber-capable AI systems should be tested and contained.
• WIRED says rogue AI hacking incidents are opening unresolved questions about liability, agency law and computer misuse rules.
• PCMag reports that Gemini's Google Docs comment features can generate summaries, replies and suggested edits for Workspace users.

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.