27 May 2026: ChatGPT Moves Into Money Management (AM)

Today’s AI news is about trust moving closer to the product. ChatGPT is asking some users to connect financial accounts, Google is pushing Gemini further into agents, Anthropic is explaining how it keeps Claude contained, and UK officials are tightening links with Australia on AI security.

OpenAI has launched a personal finance preview inside ChatGPT for Pro users in the US, with account connections handled through Plaid. According to OpenAI’s announcement, the new Finances experience lets eligible users connect bank, card, investment and loan accounts, then ask ChatGPT questions about spending, subscriptions, portfolio performance and planning. The company says ChatGPT cannot make financial transactions, and that disconnected account data is deleted from OpenAI systems within 30 days.

For UK readers, the important point is not immediate availability. The feature is a US preview. The signal is that AI assistants are becoming a front end for sensitive personal data, not just a place to draft emails or summarise documents. That makes consent, deletion controls and account level permissions more important than the novelty of a chart in a chat window. Anyone testing this kind of tool should treat it as a budgeting aid, not a substitute for regulated financial advice.

Small business owners should watch this closely because the same pattern is likely to reach bookkeeping, expenses and cash flow tools. A connected assistant could save time by spotting late payments or duplicate subscriptions, but the risk is that a confident answer can still be wrong. Cristoniq’s guide to why AI gets things wrong even when it sounds confident is the background worth reading before linking any sensitive account data to an AI system.

Google has released Gemini 3.5 Flash and used I/O 2026 to frame its next AI push around agents rather than simple chatbots. In Google’s model announcement, the company says Gemini 3.5 Flash is available through the Gemini app, AI Mode in Search, Google Antigravity, the Gemini API in AI Studio and Android Studio, and enterprise products. Google also says it is beginning trusted tester access for Gemini Spark, a personal agent designed to work in the background under user direction.

The practical shift is that Google is making agents part of its core consumer and developer stack. The benchmark figures in the announcement are vendor reported, so they should be treated as claims until independent testing catches up. What matters today is availability: developers can start testing the model in tools they already use, and everyday users will see more agentic behaviour in Search and Gemini. Cristoniq’s explainer on what an AI agent is gives useful context for why this is different from asking a chatbot one question at a time.

Anthropic has published a detailed engineering note on how it contains Claude across products such as claude.ai, Claude Code and Cowork. The Anthropic engineering post explains how the company thinks about sandboxing, permissions, service boundaries and the risk of giving an AI system access to tools that can affect real infrastructure.

This is not a flashy product launch, but it is one of the more useful reads of the week for teams adopting AI agents. As assistants move from answering questions to writing code, using browsers and touching business systems, containment becomes a core product feature. The lesson for smaller firms is simple: do not judge an AI tool only by how clever the demo looks. Ask what it can access, what it can change, how approvals work, and whether the provider has thought seriously about what happens when an agent makes a bad call.

Figma has started rolling out a design agent inside its collaborative canvas, turning AI into a participant in the workspace rather than a separate prompt box. In Figma’s launch post, the company says the agent can help generate and edit interface ideas while keeping the user inside Figma Design. The feature sits alongside Figma’s broader work on agents, design systems and developer handoff.

This matters because design tools are not just image generators. They are shared workspaces where teams make decisions, leave comments, manage components and hand work to engineers. If Figma’s agent can work inside that structure, it may be more useful than a standalone tool that produces a screen but loses the design system around it. The next test is whether teams can trust the agent with real brand rules, accessibility requirements and product constraints, not just quick mockups.

Stability AI has released Stable Audio 3, a new family of audio models that can generate longer music and sound effects, with open weights for smaller versions. According to Stability AI’s research release, the small and medium models are being released with weights and training code, while the larger model is available through hosted or paid deployment routes. TechCrunch reported that the top model can generate compositions longer than six minutes, but the company claim should still be treated as vendor reported until creators test it broadly.

The interesting part is not just song length. It is the licensing and deployment direction. Stability says the release is built for artistic experimentation and includes models that can run on consumer grade hardware. That could matter to video editors, podcasters and small studios that want quick sound sketches without sending every idea to a cloud tool. It also keeps copyright and training data questions near the surface, which is where they belong in AI music.

The UK and Australia have agreed deeper cooperation between their AI safety and security institutes as frontier systems become more capable in cyber tasks. A GOV.UK statement says the UK AI Security Institute and the Australian AI Safety Institute will share information on frontier AI capabilities, collaborate on evaluation practice and exchange research findings.

For readers, this is a reminder that AI safety is becoming operational, not theoretical. Governments are now looking at whether powerful models can help defenders but also make attacks easier. The near term impact is likely to be more evaluation, more procurement caution and more pressure on vendors to explain how their systems behave in cyber settings. Businesses using AI for code, customer data or infrastructure should expect security questions to become part of ordinary vendor due diligence.

Here is everything else worth knowing from today’s AI news.

• Google Workspace gains more voice based AI features, with Google saying Gmail, Docs and Keep are getting conversational voice features tied to its wider I/O push.
• Google Antigravity 2.0 is aimed at agent developers, with developer updates covering Managed Agents in the Gemini API and native Android support in Google AI Studio.
• Figma’s May release notes explain the agent workflow, with its help centre describing skills, Model Context Protocol support and write to canvas workflows.
• Google’s I/O collection is worth scanning, because the official roundup shows how model, app, search and developer announcements now fit into one agent strategy.

The thing to watch over the next few weeks is whether these agent features stay bounded once real users connect accounts, documents, design systems and codebases. The winners will not be the assistants that promise the most autonomy. They will be the ones that make permissions, corrections and audit trails boring enough for ordinary work.

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning.