AI Daily 13 Apr: Anthropic Finds Zero-Days, Meta Drops OSS

13 April 2026 — Anthropic has unveiled a powerful new AI built for cybersecurity and chosen not to sell it, while Meta has quietly abandoned open source with a new proprietary model heading to WhatsApp and Instagram.

The biggest AI story this week is one where the company involved is actively choosing not to release its product. Anthropic announced Claude Mythos Preview, a frontier model that has already identified thousands of software vulnerabilities — including a 17-year-old security flaw in one of the world’s most widely used operating systems. Rather than launching it commercially, Anthropic is deploying it only through Project Glasswing, a coalition that includes Apple, Google, Microsoft, Amazon, and Cisco, focused on securing critical global infrastructure.

Elsewhere, Meta has made a significant break from its long-standing values. Its new Muse Spark model is closed and proprietary, reversing years of commitment to releasing its best AI models as open source. There is also major news on funding, regulation, and tools many readers use every day: Google has made its Gemini app considerably more useful, OpenAI has closed the largest private funding round in technology history, and the UK’s financial regulator is expanding its AI testing programme.

1. Anthropic Unveils Claude Mythos Preview via Project Glasswing

An AI model that can find security vulnerabilities at a scale no human team could match is now operational — but Anthropic has decided not to release it to the public.

Claude Mythos Preview has already identified thousands of zero-day vulnerabilities (previously unknown security flaws) across every major operating system and web browser. Among them was a 17-year-old remote code execution flaw in FreeBSD, a widely used server operating system. Anthropic deployed the model through Project Glasswing, a coalition of more than 40 organisations including Apple, AWS, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, and the Linux Foundation — all focused on securing critical global software infrastructure.

The decision not to offer Mythos commercially is itself the story. Anthropic is explicitly stating that this model is powerful enough to be dangerous in the wrong hands, and is treating restricted access as the responsible choice. For anyone who uses a computer, a browser, or modern software — which covers most people — the vulnerabilities being surfaced affect systems they rely on daily. UK organisations running Linux, macOS, Windows, or Chrome infrastructure are directly affected by the flaws being uncovered, even though no UK-specific partner is named in the coalition.

2. Meta Launches Muse Spark — and Quietly Drops Open Source

Meta has built its AI identity around giving its best models away for free. That era appears to be ending.

Muse Spark, released this week as the first model from Meta Superintelligence Labs, is closed and proprietary — a deliberate departure from the Llama model series. Developed under chief AI officer Alexandr Wang (formerly of Scale AI, in which Meta has invested $14.3bn), Muse Spark is a natively multimodal model (able to understand both text and images) available in three reasoning modes: Instant, Thinking, and Contemplating. It reportedly matches the capability of the older Llama 4 Maverick while using one tenth of the compute. Meta plans to roll it out to Facebook, Instagram, WhatsApp, and Messenger in the coming weeks.

For ordinary users, the most immediate change will be a significantly more capable AI assistant appearing inside apps they already use. WhatsApp has strong UK penetration, making this directly relevant for UK readers. The commercial logic for Meta’s open-source reversal is straightforward: the company has committed an estimated $115–135bn in 2026 capital expenditure and can no longer afford to hand its best work to competitors.

3. Google Brings NotebookLM Into the Gemini App

Google has made its AI assistant considerably more useful for anyone who works with documents, research, or ongoing projects.

NotebookLM — Google’s tool for organising and querying documents using AI — is now integrated directly into the Gemini app as a feature called Notebooks. Users can group conversations, documents, PDFs, YouTube videos, and website links into persistent project workspaces inside Gemini, with everything automatically synced. The rollout began on 8 April for Google AI Ultra, Pro, and Plus subscribers on the web, with free users and additional European countries following in coming weeks.

Until now, Gemini has largely been a stateless chat interface — meaning it had no memory between sessions. This integration moves it toward a persistent, contextual AI workspace, a direction Microsoft Copilot and Apple Intelligence have both been pursuing. UK users have frequently been on delayed rollout timelines for Google features; the explicit mention of European expansion makes this one to watch in the weeks ahead.

4. Anthropic’s Revenue Hits $30bn Run Rate, Backed by Massive Compute Deal

Behind Anthropic’s safety-focused headlines is a company growing at extraordinary speed.

Anthropic disclosed that its annualised revenue run rate has surpassed $30bn, roughly 3.3 times the $9bn recorded at the end of 2025. More than 1,000 enterprise customers are now spending over $1m per year with the company. To support that growth, Broadcom confirmed a long-term agreement to supply Anthropic with 3.5 gigawatts of Google TPU (Tensor Processing Unit — a type of custom AI chip) compute capacity from 2027, building on a 1 gigawatt deal already in place for 2026. Mizuho analysts estimate Broadcom will record $21bn in AI revenue from Anthropic in 2026 alone, rising to $42bn in 2027.

The compute infrastructure being assembled reinforces concerns about concentration in the AI industry. A small number of companies — Google, Broadcom, Amazon, and NVIDIA — now supply the physical foundation on which the world’s most advanced AI systems run. Anthropic closed a $30bn Series G funding round earlier this year at a $380bn valuation.

5. OpenAI Closes $122bn Round — Amazon and NVIDIA Lead

The largest private funding round in technology history is now complete.

OpenAI closed a $122bn funding round led by Amazon ($50bn), NVIDIA ($30bn), and SoftBank ($30bn), at a post-money valuation of $852bn. SoftBank’s second $10bn tranche arrived on 1 April 2026. As part of the Amazon deal, OpenAI will expand its AWS partnership with an additional $100bn in committed compute services, build a stateful runtime environment on Amazon Bedrock, and develop custom models for Amazon consumer products. OpenAI’s annualised revenue has now passed $25bn, with enterprise customers accounting for more than 40% of total revenue.

The deal effectively ends Microsoft’s near-exclusive hold over OpenAI’s infrastructure. An OpenAI IPO is being explored for late 2026 at a near-trillion-dollar valuation — which would represent one of the most significant public listings in years, and would make OpenAI investable for UK retail and institutional investors for the first time.

6. OpenAI Acquires Tech Talk Show TBPN

OpenAI has made its first-ever media acquisition — a move that has raised immediate questions about editorial independence.

TBPN (Technology Business Programming Network), a daily founder-focused tech talk show hosted by John Coogan and Jordi Hays, was acquired by OpenAI for a price reported to be in the “low hundreds of millions.” The show generated around $5m in advertising revenue in 2025 and is projected to reach $30m in 2026. TBPN will reportedly maintain editorial independence but will report to OpenAI’s chief political officer, Chris Lehane — a structuring detail that drew particular scrutiny, given the obvious potential for conflicts of interest when an AI company owns a media outlet covering AI.

An AI lab owning a media outlet is new territory. OpenAI framed the deal as creating “a space for constructive conversation about AI and its changes.” TBPN has a UK listener base, and the media governance questions raised by this acquisition have relevance to ongoing UK debates about platform power and press regulation.

7. FCA Opens Second Cohort of AI Live Testing for UK Financial Services

The UK’s financial regulator is moving faster on AI in practice than its public statements might suggest.

The Financial Conduct Authority launched the second cohort of its AI Live Testing programme in April 2026, offering financial services firms a controlled environment to test AI models in real market conditions, with regulatory oversight and synthetic data support from technical partner Advai. On 1 April 2026, the FCA and the Prudential Regulation Authority jointly reaffirmed a technology-neutral, principles-based approach to AI — meaning no AI-specific rulebook is being written for now.

For fintech and financial services companies operating in UK markets, the live testing pathway is one of the most concrete regulatory options available. The FCA’s 2026/27 work programme also confirms a new Supercharged Sandbox cohort. The Treasury Committee warned in January 2026 that the regulators’ current stance “risks serious harm to consumers” — a debate that is likely to intensify as the August 2026 EU AI Act deadline approaches.

At a glance

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

• Google Gemma 4 released as open-weight model family — Four variants available under an Apache 2.0 licence, with major benchmark improvements over Gemma 3 (AIME 2026 performance up from 20.8% to 89.2%), targeting agentic and edge deployment. Google DeepMind
• Gemini 3.1 Flash-Lite now in preview — Google’s most cost-efficient Gemini model at $0.25 per million input tokens, 2.5 times faster time-to-first-token than 2.5 Flash, with adjustable reasoning levels. Available on Vertex AI and the Gemini API. Google Blog
• OpenAI enterprise revenue surpasses 40% of total; IPO preparations reportedly under way — Annualised revenue past $25bn; IPO timeline targeting late 2026. OpenAI
• NVIDIA Physical AI Data Factory Blueprint announced — Open reference architecture for automated generation and evaluation of physical AI training data, targeting robotics and autonomous vehicle development. NVIDIA Newsroom
• EU Digital Omnibus progressing through Parliament — EU proposals to simplify the AI Act and GDPR are under examination; the main high-risk AI compliance deadline falls on 2 August 2026. artificialintelligenceact.eu
• BoE and PRA restate technology-neutral AI approach — Published 1 April 2026 in response to DSIT’s strategic direction; no AI-specific rules planned, existing frameworks considered sufficient for now. Inside Global Tech
• UK AI Bill remains pre-introduction; King’s Speech expected May 2026 — Government focus has shifted toward AI growth over safety; the Bill is not expected to be tabled until at least the next King’s Speech. UK AI Regulation analysis
• n8n 2.0 with sandboxed code execution and persistent agent memory now widely deployed — n8n’s 2.0 platform, launched in January 2026, continues to gain traction as a capable AI workflow automation tool for businesses building their own AI pipelines without coding expertise. Automation Atlas
• Windsurf SWE-1.5 coding model scores 40% on SWE-Bench — Matches Claude Sonnet accuracy at 14 times the tokens-per-second speed; Arena Mode for side-by-side model comparison now available with thousands of developer votes logged. Roborhythms
• OpenAI GPT-5.4 is current flagship; GPT-5.2 Thinking being retired on 6 June 2026 — GPT-5.4 launched 5 March 2026, combining frontier coding with reasoning. GPT-5.2 Thinking is being removed from the model picker. OpenAI Help Centre
• International AI Safety Report 2026 published in February — Led by Turing Award winner Yoshua Bengio and backed by more than 30 countries, the report finds frontier models performing at or above human expert level on standardised evaluations, and identifies an evaluation gap in pre-deployment testing. internationalaisafetyreport.org

UK and regulatory update

The most significant UK development right now is the FCA’s second AI Live Testing cohort, now under way in April 2026. This offers financial services firms a supervised, real-world testing pathway that very few comparable frameworks provide. The Bank of England and PRA reaffirmed on 1 April 2026 that existing regulatory frameworks are sufficient for AI oversight, though the Treasury Committee warned in January that this stance risks harm to consumers and the financial system. The UK AI Bill remains pre-Parliamentary, with the earliest opportunity for introduction being a King’s Speech in May 2026. For firms with EU operations, the August 2026 deadline for high-risk AI Act provisions is now under 16 weeks away, and the EU’s Digital Omnibus proposals to simplify the Act and GDPR are moving through Parliament.

What to watch

• UK AI Bill timeline: Any announcement from Downing Street about the Parliamentary programme for spring 2026 will signal when the AI Bill is likely to be introduced. Watch DSIT and No. 10 channels.
• Claude Mythos access expansion: Anthropic has indicated the goal is wider Mythos-class deployment. Watch for updates on whether Project Glasswing expands beyond its initial 40+ partner organisations, and whether any disclosed vulnerabilities receive public CVE listings.
• Muse Spark rollout to WhatsApp and Instagram: Meta indicated deployment to UK-facing platforms in the coming weeks. The WhatsApp launch will be the most significant consumer moment for UK users.
• OpenAI IPO signals: With SoftBank’s April tranche confirmed and valuation nearing $900bn, any regulatory filing or executive statement about listing timeline would be significant.
• EU AI Act compliance clock: Now under 16 weeks to the 2 August 2026 high-risk provisions deadline. Watch for guidance from the European AI Office and signals from national enforcement authorities.

What this means for you

The most practical takeaway from today is that AI is now actively improving the security of the software you use every day. The Anthropic Project Glasswing coalition is finding and patching long-standing vulnerabilities in Windows, macOS, Linux, and Chrome — software updates arriving over the coming months are worth installing promptly. If you use WhatsApp or Instagram, a noticeably more capable AI assistant is heading to those apps in the coming weeks as Meta rolls out Muse Spark. And if you are a small business owner in UK financial services, the FCA’s AI Live Testing programme is worth examining: it offers a supervised, real-world pathway to deploying AI tools that very few comparable frameworks currently provide.

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.

Note: 21 sources in today’s research sweep were not directly accessible and returned no relevant results. All stories in this post are sourced from accessible, attributed publications.

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