AI Daily Brief: 29 March 2026

29 March 2026

This Sunday's AI landscape is dominated by two major stories that have significant implications for UK businesses: a leaked Anthropic model that could redefine the frontier of AI capability, and fresh evidence that the UK government's much-announced AI ambitions are not translating into action. Add a dramatic pivot at OpenAI, China's chip breakthrough, and an escalating deepfake crisis, and there is plenty to digest before the week begins.

Anthropic's "Claude Mythos" Leaks — and It Rewrites the AI Capability Map

On 27 March 2026, internal Anthropic documents including draft announcement posts and nearly 3,000 unpublished assets were accidentally exposed through a misconfigured, publicly searchable database. The documents revealed a new model codenamed "Claude Mythos" — also considered under the name "Capybara" — that represents a category above the existing Opus line.

Anthropic confirmed to Fortune that the model is real and in testing. A spokesperson described it as "a step change and the most capable we have built to date," with "meaningful advances in reasoning, coding, and cybersecurity." According to the leaked drafts, it achieves "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity" compared to Claude Opus 4.6.

The most striking element is Anthropic's planned release strategy. The company describes the model as "currently far ahead of any other AI model in cyber capabilities," capable of exploiting vulnerabilities "in ways that far outpace the efforts of defenders." The rollout will begin with a small, vetted group of security researchers before any public availability.

Our take: For UK businesses, this signals two things. First, the next frontier of AI capability is closer than the current public model landscape suggests. Second, Anthropic's willingness to slow-walk a major product for safety reasons is itself significant — and will likely inform how enterprise clients are brought on board. Watch for a controlled, compliance-aware release that favours established enterprise relationships over open access.

The UK Government's OpenAI Partnership: Eight Months, Zero Trials

A Freedom of Information request submitted to the Department for Science, Innovation and Technology (DSIT) has produced a damaging result: the government confirmed it has "not undertaken any trials under the memorandum of understanding for OpenAI." The MOU was signed eight months ago with considerable fanfare, promising to harness AI to "address society's greatest challenges." So far, the only concrete deployment of advanced AI in government is ChatGPT in the Ministry of Justice.

A separate investigation found that Nscale, which was publicly committed to building the UK's largest AI supercomputer by the end of 2026 using Nvidia GPUs, has been misrepresenting progress on the project and is very unlikely to complete it on time.

Our take: The private sector cannot afford to wait for government to set the pace on AI adoption. The evidence increasingly suggests that UK public sector AI ambition is running years ahead of execution. For businesses, this is both a risk (regulatory clarity is slow in coming) and an opportunity (competitive advantage from early, practical AI adoption is real and durable).

UK AI Copyright Reform Delayed — Again

The UK government published its report on copyright and AI training data on 18 March 2026 — and confirmed there will be no immediate reform. Transparency requirements, licensing frameworks, and deepfake regulation are all being moved to further consultation rounds. The report was published under sections 135 and 136 of the Data (Use and Access) Act 2025 and was long-awaited by both creative industries and AI developers.

The House of Lords Communications and Digital Committee had previously warned that the UK must choose between two AI futures. That choice, it appears, has been deferred. No new central AI regulator is proposed. Statutory powers for oversight remain on hold.

Our take: UK businesses using AI-generated content or training models on proprietary data face continued uncertainty over their legal position. The practical advice remains: document your data provenance, use commercially licensed training data where possible, and treat transparency about AI use in content as a reputational baseline rather than a legal requirement — because legal requirements are still being worked out.

OpenAI Scraps Sora, Kills Disney Deal, and Raises $10 Billion in One Day

In a single extraordinary Tuesday, OpenAI announced it would shut down Sora, its video-generation product; wind down a $1 billion partnership with Disney; shuffle executive roles; and close an additional $10 billion funding round, bringing its total raise to more than $120 billion.

The reasoning is blunt: Sora consumed enormous compute without generating commensurate revenue, and had fallen behind competing video models. Fidji Simo, now CEO of AGI Deployment after being moved from her applications role, reportedly told staff: "We cannot miss this moment because we are distracted by side quests. We really have to nail productivity in general and particularly productivity on the business front."

Our take: OpenAI is undergoing a decisive pivot from experimental showcase products toward profitable enterprise Tools. The $120 billion funding total reflects investor appetite for AI infrastructure, but also the enormous burn rate that comes with it. For UK businesses evaluating ChatGPT and OpenAI's enterprise products, this focus on productivity and commercial return is actually good news — it means OpenAI's priorities are increasingly aligned with yours.

Huawei AI Chip Wins ByteDance and Alibaba Orders — A Geopolitical Turning Point

Reuters reports that Huawei's new AI chip, designed to directly challenge Nvidia in the Chinese market, has completed successful customer testing with ByteDance and Alibaba — both of which now plan to place orders. The chip represents the most significant milestone yet in China's push for semiconductor self-sufficiency.

US export controls on Nvidia chips have been in place for several years, but rather than slowing Chinese AI development, the evidence suggests they have accelerated domestic alternatives. If ByteDance and Alibaba commit fully to Huawei silicon, it marks a structural shift in how China's AI industry powers itself.

Our take: For UK businesses operating globally or in supply chains with Chinese technology components, this is a development worth tracking. It adds another dimension to the already complex question of AI supply chain resilience. It also has implications for Nvidia's revenue projections and, by extension, the broader AI infrastructure investment thesis.

AI Deepfakes Are Now a Mainstream Propaganda and Revenue Tool

The Guardian has published a detailed investigation into how AI-generated military imagery and video deepfakes have become both propaganda tools and revenue-generating content. Faceless social media accounts are producing and monetising AI-generated images — including women in military contexts — as part of content farms serving global audiences. Meanwhile, deepfake war videos depicting events in the Middle East are circulating widely despite having no basis in reality.

Researchers describe the "liar's dividend" effect: the mere existence of deepfakes means genuine footage can be credibly denied, and false footage can be made to appear plausible. This applies across news, politics, and commercial content.

Our take: For UK businesses, the deepfake problem is not abstract. Brand impersonation, executive voice cloning for fraud, and AI-generated testimonials are already operational threats. The practical response is to build verification habits into your media consumption and communications processes — and to assume that any attention-grabbing video or audio content may require source confirmation before acting on it.

Suno v5.5: Voice Cloning Comes to AI Music

Suno released version 5.5 of its AI music platform on 26 March, adding three new features: Voices (voice cloning for personalised output), Custom Models (fine-tuned personal music AI), and My Taste (preference-based generation). Users can now produce songs sung in their own voice, marking a significant personalisation step for the platform.

The release arrived simultaneously with Google DeepMind's Lyria 3 Pro, which promises improved instrumental rendering and dynamic control. The AI music space is becoming increasingly crowded and capable.

Our take: AI-generated music is moving from novelty to production tool at pace. For marketing teams, this opens up cost-effective options for branded audio content, video soundtracks, and social media clips. The voice cloning element also raises fresh intellectual property questions — particularly for talent and music rights holders.

Quick Hits

• AI productivity gains are real and measurable: A CTO writing for VentureBeat reports that his engineering organisation achieved 170% throughput at 80% headcount after going "AI-first" — a 2x improvement backed by six months of objective data. For UK business leaders still debating AI adoption, this is worth reading in full.
• TikTok is not labelling AI-generated ads: The Verge reports that AI-generated advertising content is appearing on TikTok without the disclosures required by the platform's own policies — including from major brands like Samsung, which does disclose AI use on other platforms. This is a transparency and brand risk issue as much as a regulatory one.
• UK public data "not yet usable" for AI: ODI testing of the government's National Data Library has revealed serious gaps in data quality and accessibility, undermining ambitions to use public sector data for AI development.

Frequently Asked Questions

What is Anthropic's Claude Mythos and when will it be released?

Claude Mythos is Anthropic's next-generation AI model, accidentally revealed via a database misconfiguration in March 2026. Anthropic has confirmed it is real and describes it as a step change that outperforms all previous Claude models on coding, reasoning, and cybersecurity. Due to its advanced cyber capabilities, Anthropic plans a deliberately slow, security-focused rollout beginning with a small group of vetted researchers. No public release date has been announced.

Why has the UK government not used its OpenAI partnership?

A Freedom of Information request revealed that the UK's Department for Science, Innovation and Technology had not undertaken any trials under its memorandum of understanding with OpenAI, despite the agreement being signed eight months before. The MOU was promoted as a landmark step toward AI-led public service reform. Analysts suggest the gap reflects slow procurement processes and unclear governance frameworks rather than lack of intent.

What does the UK government's copyright and AI report mean for businesses?

The UK government's March 2026 report on copyright and AI training data confirmed that no immediate legislative reform will be introduced. Key questions around transparency, licensing for AI training, and deepfake regulation have been deferred to further consultation. For businesses, the legal grey zone around AI-generated content and training data persists. The practical advice is to document data provenance, use commercially licensed training datasets, and treat transparency about AI use as a reputational best practice.