AI Daily 12 Apr: Anthropic Pulls Claude Mythos Over Risks

12 April 2026 — Anthropic has refused to publicly release its most capable model yet because it can help attackers find holes in everyday software, and the fallout is already reaching UK banks.

It is a rare day when the biggest story in AI is a model that is not shipping. Anthropic’s decision to lock down Claude Mythos Preview has forced regulators, enterprises and rival labs into an uncomfortable conversation about what frontier models can now do, and what should happen when a lab decides the risks outweigh the revenue. Alongside that, California has taken a first swing at Grok over deepfakes, an open-source agent project has leapfrogged one of its better-known rivals, and Canva has quietly shipped what may be the first true prompt-to-on-brand-asset workflow for small businesses.

Here is what matters from today’s AI news, and why it should matter to you.

1. Anthropic withholds Claude Mythos Preview over cyber-offence risk

Anthropic confirmed on Friday that Claude Mythos Preview, its newest frontier model, will not be released to the public. According to reporting from Bloomberg, CSMonitor and CP24, the model can identify exploitable vulnerabilities in nearly every major operating system and web browser from relatively simple prompts. Instead of a normal launch, Anthropic is distributing it under an invitation-only research preview called Project Glasswing, with access limited to Amazon, Google, Apple, Microsoft, Nvidia and Cisco for defensive security work.

Some of Anthropic’s own developers have publicly said the model is too dangerous to ship, which is itself a first for a major US lab. A vulnerability, in plain terms, is a weakness in software that an attacker can exploit to break in or steal data.

Why it matters for you. If you run a small business or use cloud tools for work, this is not an abstract story. It means the calculus around how quickly you install software updates just changed. The Bank of England is convening UK banks next week to discuss exposure, mirroring moves by US and Canadian regulators. Expect a push from banks, cloud providers and IT teams to tighten patch cycles and review third-party software risk.

2. Anthropic’s revenue rockets to $30 billion as it secures Google and Broadcom compute

Anthropic has extended its compute agreement with Google Cloud and Broadcom to secure more TPU capacity, the custom AI chips Google designs for training and running large models. TechCrunch reports the company’s run-rate revenue has reached about $30 billion, up from roughly $9 billion at the end of 2025. Anthropic also raised a $30 billion Series G funding round in February.

The expanded deal is tied directly to the compute footprint needed for Mythos-class research and the next generation of Claude. In plain English, training the most capable models now requires bespoke silicon deals, not just pulling GPUs off a shelf.

For businesses choosing between AI providers, this is confirmation that Anthropic has closed a meaningful share of the enterprise gap with OpenAI. Pricing and model quality should both benefit from the increased competition at the top of the market.

3. Nous Research ships Hermes Agent v0.8.0 and overtakes OpenClaw on GitHub

Nous Research released Hermes Agent v0.8.0 on 8 April. The update adds a persistent memory layer, credential rotation, an anti-detection browser called Camofox, inline diffs and a run of stability fixes from the earlier v0.7.0 release. The project has added more than 6,400 GitHub stars in a single day and now sits above 47,000, according to coverage from PANews, openPR and 36Kr. Developers are increasingly choosing it over OpenClaw for self-hosted agent stacks, and OpenClaw has not shipped a comparable update in the last 24 hours.

An AI agent, for the uninitiated, is a program that uses a language model to plan and carry out multi-step tasks by itself, such as researching a topic or booking travel. Hermes Agent is becoming the default open-source option for developers who want that kind of autonomy without handing their data to a hosted service.

If you are a small business evaluating automation, the practical takeaway is that the self-hosted agent scene has matured enough that you no longer have to rely on a single proprietary platform. That competition will show up as better features and lower prices across the board.

4. California orders xAI to stop Grok generating deepfakes of real people

California Attorney General Rob Bonta has issued a formal demand to xAI requiring it to immediately stop its Grok model from producing non-consensual deepfake imagery of real people. VentureBeat and TechCrunch report the order follows weeks of complaints about Grok 4.20’s image generation pipeline, which had been marketed for its loose content filters. A deepfake is a synthetic photo, video or audio clip made to look or sound like a real person. xAI has not yet responded publicly.

This is one of the first state-level enforcement actions against a frontier AI lab over synthetic media. It is likely to force a compliance rewrite of Grok’s image tools before the planned Grok 5 launch.

UK observers will be watching closely. The Information Commissioner’s Office has previously flagged non-consensual deepfakes as a priority enforcement area, and California’s move gives UK regulators a template to follow.

5. Mistral releases Voxtral open-weights voice model and joins NVIDIA’s Nemotron Coalition

Mistral has published Voxtral TTS on Hugging Face under a non-commercial open licence. The model offers zero-shot voice cloning, multilingual coverage and real-time streaming. In the same news cycle, NVIDIA announced the Nemotron Coalition with Mistral, Perplexity, Cursor, Black Forest Labs, LangChain, Reflection AI, Sarvam and Thinking Machines Lab as founding members committed to shared research, data and compute for open frontier models.

Voice cloning means the model can learn to speak in a specific person’s voice from a short sample. The fact that this is now available as open weights lowers the bar significantly, which is useful for creators and accessibility work but also raises the same deepfake concerns the California case is about.

The Nemotron Coalition is a direct challenge to the closed-lab axis that has dominated the last year. For businesses, more open competition should mean more choice over where your data lives.

6. n8n 2.0 brings native AI agents and persistent memory to open-source automation

n8n’s 2.0 line has added a native AI Agent Tool Node for multi-agent orchestration, LangChain integration, sandboxed code execution and persistent agent memory. Zapier this week pushed a matching Copilot natural-language builder update, and Make.com extended its Grid governance layer.

In plain English, n8n is a self-hosted tool for stringing apps and AI together into automated workflows, the open-source equivalent of Zapier or Make. The 2.0 release closes most of the gap with hosted rivals.

If you run a small business, the automation platforms are quietly becoming the main surface for practical AI adoption. The choice between self-hosted and hosted now comes down to data preferences and in-house skills rather than features. n8n is popular with UK indie developers and agencies, so it is an easy place to start if you want to keep data on your own servers.

7. Canva adds on-brand design generation via its Claude connector

Canva has expanded its Claude connector so that Brand Kit colours, fonts and voice are applied automatically when designs are generated from a Claude conversation. You can now prompt Claude for a social post or pitch slide and get an on-brand Canva draft back in a single step. Futurum Group and the Canva newsroom covered the update, which builds on Canva’s earlier Magic Studio work.

For marketing teams and small businesses, this is arguably the first true end-to-end prompt-to-on-brand-asset workflow. You describe what you want, and what comes out already uses your colours, your fonts and your voice, with no human designer in the loop for first drafts.

It will not replace a designer for nuanced work. It will take a lot of the repetitive social and internal-deck grind off their plates.

At a glance

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

• GPT-5.3 Instant Mini replaces GPT-5 Instant Mini in ChatGPT — OpenAI has rolled out GPT-5.3 Instant Mini as the default fallback after rate limits. Source: OpenAI Help Center.
• Google DeepMind Gemini 3.1 Flash-Lite pricing drops to $0.25 per million input tokens with 2.5x faster responses. Source: Google Developers Blog.
• Gemma 4 on-device agent models widely available for offline multi-step agents. Source: Google Developers Blog.
• Meta releases Muse Spark closed-source model, ranked 4th on the Artificial Analysis Intelligence Index v4.0. Source: CNBC.
• Microsoft ships three new foundational models for text, voice and image as part of its in-house stack. Source: TechCrunch.
• Cursor applies real-time reinforcement learning to Composer, shipping improved versions as often as every five hours. Source: Releasebot Cursor updates.
• Perplexity annual recurring revenue reaches $305m, up 50%, though fresh lawsuits allege copyright infringement. Source: blog.mean.ceo and Crescendo AI.
• Notion desktop gains voice input and shareable AI chat links, plus image generation inside pages. Source: Notion releases.
• Spirit AI (China) raises about $420m across two rounds, backed by Shunwei Capital and Yunfeng Fund. Source: The AI Insider.
• Cyberhill Partners secures up to $11m from Baleon Capital to scale enterprise AI consulting. Source: Fintech Global.
• Q1 2026 global venture funding hit a record $300 billion across 6,000 startups, driven by AI compute deals. Source: Crunchbase News.
• European Commission AI and GDPR simplification drive criticised by Amnesty International as a rollback of rights. Source: Amnesty International.
• UK FCA and PRA confirm a principles-based AI approach for 2026, reaffirming no AI-specific rulebook. Source: Resultsense.
• Boston Dynamics and Google DeepMind robotics partnership continues to generate follow-on coverage as humanoid foundation models take shape. Source: TechCrunch and Boston Dynamics.
• Indian startup Rocket launches AI-generated McKinsey-style consulting reports at a fraction of consulting prices. Source: TechCrunch.

UK and regulatory update

The most significant UK angle today is the Bank of England’s decision to convene UK banks next week to discuss the risks posed by Anthropic’s withheld Claude Mythos Preview. This is a rare pre-emptive supervisory move tied directly to a frontier model, and it mirrors parallel action in the US and Canada. Separately, the FCA, PRA and Bank of England have reaffirmed their principles-based, technology-neutral approach to AI in 2026 rather than pursuing AI-specific rules, and the FCA’s Mills Review of AI in retail financial services is still underway. In Europe, the Commission is pushing a simplification agenda for the AI Act and GDPR that Amnesty International says will weaken rights, ahead of the 2 August full-application milestone. The overall picture: the UK is responding to Mythos but continues to hold a lighter regulatory line than the EU.

What to watch

• Follow-up from the Bank of England’s Mythos briefing to UK banks and any joint statement with the FCA or NCSC.
• A formal xAI response to California’s Grok deepfake order.
• A possible first public Claude Mythos technical paper or safety card.
• Grok 5 timing signals from xAI ahead of its Q2 launch window.
• Further Nemotron Coalition detail, including joint benchmark and dataset plans from Mistral, Cursor, Perplexity and Black Forest Labs.

What this means for you

The single most practical takeaway is this. If you run a small business or manage IT for one, treat the next two weeks as a patching window. The Mythos story tells us that models now exist that can find software holes faster than most defenders can fix them, and those models are in the hands of a small number of major platforms. Update your operating systems, browsers and core business apps as soon as prompts appear. If you use Grok for image work, assume its image tools are about to change and plan around it. And if you have been putting off trying a self-hosted automation tool like n8n, today is a reasonable day to spin one up.

Disclaimer

This is an automated 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 evening.