Good morning from the wAIve team in London.

Today, Tuesday 10 November 2026: A day of bold claims and awkward questions. Anthropic says its newest models now write more than nine-tenths of its own code, and Australia wants AI companies to own up, and quickly, when their systems go rogue. Meanwhile, Nvidia's Jensen Huang is discovering that being the President's favourite chief executive doesn't keep the FBI away.

  BREAKING  

Jensen Huang's crisis deepened overnight. Trump confirmed the FBI is investigating the scandal and shareholder lawsuits are expected, with experts warning that Nvidia's earlier characterisation of the relationship as "normal" could prove costly. A Yale governance expert put Huang's chances of keeping his job at roughly even. Huang has withdrawn from at least two industry events as his board conducts its independent review.

THE PLAY: Huang was the industry's most effective lobbyist for loosening US chip export restrictions. With a Democrat-controlled Congress preparing to ramp up scrutiny of chip sales to China, the politics has changed.

  INDUSTRY  

Anthropic has released Claude Fable 5.2 and Mythos 5.2, described as its most capable models to date. According to the company, they solve open physics problems, demonstrate advanced cybersecurity capabilities, and write more than nine-tenths of the code changes merged into Anthropic's own codebase. The models come in two versions: one open to all users and one restricted to verified organisations. Anthropic says it has deliberately limited any capabilities that would allow the AI to accelerate its own development.

THE PLAY: The launch comes days before the company is expected to set its IPO valuation range.

  POLICY  

Australia's government has announced mandatory reporting requirements for AI companies. The move follows an OpenAI data breach in June. Prime Minister Anthony Albanese confirmed changes to criminal law so that companies cannot argue in future that they were unaware their systems caused a breach. The new framework requires AI developers to maintain security controls across all development stages, notify the government rapidly of incidents, allow independent model evaluation, and publish safety and governance frameworks.

THE PLAY: This is the strongest AI incident-reporting legislation from any non-EU government. The question is no longer whether mandatory reporting will arrive, but when. Treat as a marker for where regulation is heading globally.

WhatsApp messages obtained by the Guardian reveal that Palantir lobbied a UK health minister to extend its AI platform into GP data, beyond the scope of its existing £330 million contract with NHS England. The company argued that "the logic of joining up primary and secondary care data is unarguable" and described its software as ideal for the role. GPs and patient advocates have strongly opposed the move.

Why it matters: It's not an uncommon pattern in public-sector AI contracts: using informal channels to expand scope.

Google has now launched four custom artificial intelligence chips in a single year to serve an enterprise customer base of nearly 100 clients. The new products are the TPU 8tx and TPU 8i Max. The rollout directly challenges Nvidia's market dominance, with Bloomberg Intelligence estimating Google's direct TPU sales will skyrocket from $12 billion in 2026 to over $135 billion by 2028, excluding additional revenue from its cloud rental business.

Why it matters: This puts Google at the centre of a booming AI chip market that analysts project will expand into a $250 billion-plus market opportunity as tech giants and enterprises actively seek alternatives to Nvidia.

Disney and Adobe announced a partnership to build custom AI tools for Disney's production pipeline using Adobe's Firefly platform. The arrangement is designed to give Disney brand-safe generative models trained on approved content.

Why it matters: Disney's approach is a template for how large consumer brands can adopt generative AI without accepting the liability from general-purpose public models.

Closing prices, Monday 9 November (US markets). Not investment advice.

Medium's new writing app integrates AI assistance directly into the drafting process, offering suggestions, structure, and editing support as you write rather than generating a finished document from a prompt.

ByteDance's new image model supports stable multi-pass editing across up to 100 iterations without colour drift and generates images with transparent backgrounds suited to production use.

THE PLAY: Use Seedream 5.5 to generate three alternative visual concepts for an upcoming campaign or product launch and share them with your creative team as a brief-sharpening exercise rather than a finished output.

A string of incidents in recent weeks has raised code-red concerns around AI-assisted cyberattacks.

Blockchain analytics firm Chainalysis confirmed that attackers used AI agent-assisted techniques to breach its network, stealing credentials and session tokens for its entire customer base, which includes the FBI, US Immigration and Customs Enforcement, Barclays, and Microsoft.

The speed and scalability of attacks: Human attackers are expensive, slow, and detectable through behaviour patterns. AI agents can test vulnerabilities and deploy at a pace no human team can match.

In the Chainalysis breach the attacker used AI agents not to write malware but to navigate internal systems and extract data faster than human defenders could respond. The fridge hacks required no sophistication at the attack level, only the ability to scan thousands of internet-connected industrial controllers.

The escalation requires a board-level response, not just an IT department response.

Takeaway: AI-assisted attacks are not a future risk. They are happening now, at scale, against well-resourced organisations. Boards that have not yet integrated AI-specific threat modelling into their risk framework are operating with an incomplete picture.

Before asking AI for a recommendation on a difficult decision, try talking the problem through with it as though you were explaining it to a trusted colleague who knows nothing about your business. Most executives go straight to asking for an answer. The better approach is to use AI as a thinking partner first, surfacing assumptions you have not yet examined. This technique is sometimes called rubber ducking, after the programmer's habit of explaining code problems to an inanimate object in order to solve them.

It works because externalising your reasoning to a patient, non-judgemental listener forces you to articulate what you actually think, not what you think you should think. The AI's questions reveal gaps you would otherwise carry into a decision unchallenged.

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