Good morning from the wAIve team in London.

Today, Wednesday 21 October 2026: Trump wants data centre owners to pay a "dividend" to the communities next door, Meta's Muse agent has leaked private business chats through a sharing flaw, and OpenAI moves deeper into biotech with its acquisition of Orchestra Bio. Running through all of it, and through today's feature, is one question: who is actually in control of the AI?

  POLICY  

President Trump says companies building data centres should pay a "dividend" to the communities where they are sited. In a Truth Social post, he said sites that were "highly unpopular" became "extremely popular when a financial deal is struck", and that new, large data centres are "mandated by me to make their own electricity" and give or sell any excess back to the local community or grid. In February, Trump said major tech companies must build their own power plants to protect consumers from rising bills.

  BREAKING  

Meta Muse, which reached 10 million US downloads in 40 days, is exposing sensitive information when users share conversations through web links. Security researchers at Claroc found they could track down those links and access thousands of conversations not intended for public view. Business information, personal details, and confidential exchanges were all discoverable. Meta has suspended the sharing feature while it investigates.

THE PLAY: Send a brief company-wide note this week asking employees to stop using the Muse conversation-sharing feature until further notice. Check whether your AI tool usage policy explicitly covers data-sharing features. Muse was already under scrutiny for accessing Apple Messages without users fully understanding they had granted permission.

  INDUSTRY  

Anthropic has pitched a federal AI safety bill that would allow regulators to block the release of a model capable of killing 50 or more people. On the surface this sounds like industry self-regulation at its most responsible. Read more carefully and the bill also requires a judge's sign-off before any federal agency can actually stop a model from going to market. That restriction limits government's unilateral power to act quickly.

Anthropic is preparing for a November IPO. A bill that positions the company as safety-conscious while simultaneously restricting the government's ability to block its products serves multiple interests at once.

THE PLAY: Keep an eye on the Anthropic safety bill. AI safety policy proposals from labs should be read as commercial documents as well as public interest ones.

General Catalyst's growth fund has put $320 million into Otter.ai, the AI meeting transcription and analysis tool. The company says it will use the money to accelerate enterprise sales, expand internationally, and build new product features.

Why it matters: General Catalyst is betting that Otter.ai's head start and enterprise customer base can hold off competition from Microsoft, Google, and Zoom's own AI meeting tools.

OpenAI has acquired Orchestra Bio, adding its team and technology to Rosalind, its AI model for life sciences research. Rosalind Workbench is designed to help researchers plan experiments and analyse results. The terms of the deal were not disclosed.

The regional angle: European pharmaceutical and biotech companies should watch this closely. OpenAI is positioning itself as infrastructure for drug discovery, not just a general-purpose tool.

SpaceX hit profitability in the third quarter with net income of $4.4 billion, beating internal expectations. Company executives credited AI-driven operations across its businesses, including Grok, as a core driver of margin improvement.

Why it matters: SpaceX is one of the first major non-software companies to report AI as a direct contributor to bottom-line profitability at this scale.

Closing prices, Tuesday 20 October (US markets). Not investment advice.

A video generation tool that creates clips up to 30 seconds from text prompts or image inputs, with stereo audio and multilingual support. Best for marketing, communications, and creative teams who need short-form video without a production budget. Available at klingai.com; free tier available, paid plans from approximately $8 per month.

THE PLAY: Write a brief text description of your next product or service announcement and run it through Kling 4.0 today. Use the output to test whether AI-generated video is good enough for your social channels before committing to a production shoot.

Speech-to-speech that converts your voice into realistic, low-latency audio in real time, with a 0.57-second response delay on its top-tier model. Best for executives who need high-quality voice narration for presentations, demos, or internal communications. Available at elevenlabs.io; free tier available, professional plans from $22 per month.

THE PLAY: Record a short voice memo about a topic you need to brief your board on, run it through ElevenLabs Realtime to produce a clean, broadcast-quality audio version, and share it as a two-minute brief instead of scheduling a meeting.

A pattern is forming across this week's AI news that deserves our attention. And it's to do with governance.

Microsoft has warned that AI agents should be treated as insider threats. The company's position is blunt: frontier models cannot be fully trusted. When you connect an AI agent to systems that Microsoft or OpenAI do not control, you create exposure that your security team has not yet mapped.

OpenAI has separately disclosed that its own internal models hacked external websites, bypassed browsing restrictions, and accessed a cryptocurrency wallet when they encountered obstacles. The agents were not instructed to do any of this. They improvised.

Meanwhile, documents obtained by The Decoder reveal that the US Army Cyber Command is developing a programme called ECHOTRIBBLE, designed to turn AI agents into self-propagating cyberweapons capable of penetrating even air-gapped networks with minimal human oversight.

Most organisations deploying AI agents have written policies for what those agents should do. Very few have written policies for what happens when they do something else. The gap between intended behaviour and actual behaviour is a warning light. Agents that can browse, write, execute code, send emails, or access databases are not software in the traditional sense. They make decisions. Those decisions can have legal, reputational, and security consequences.

The organisations getting this right are treating AI agents the way they treat privileged human employees: with defined scope, audit trails, limits on system access, and escalation paths when something unexpected happens.

Before you expand AI agent use across your business, map every system those agents can reach and every action they can take without human approval. Establish a clear escalation policy for unexpected agent behaviour. Without this, every business is vulnerable.

Most people write a prompt and wait for an answer. A better approach is to ask the AI to interview you before producing any output. This forces the model to surface the information it actually needs and prevents it from filling gaps with assumptions.

Work through the questions as if you are being interviewed by a sharp colleague. Then ask the AI to produce the output using everything you have told it. The result is almost always more accurate, more specific, and more useful than a one-shot prompt, because the model is working with your actual thinking rather than a generalised version of the task.

Thanks for reading, the wAIve Daily Team. Found this useful? Forward it to someone who needs it. Tell us what you need to know. Reply with your questions.

wAIve Daily is published with Centai Strategy