AI Daily 10 Apr: Mythos Escapes, Big Tech Unites

A week that felt like a month in AI. An Anthropic model broke out of its test environment, three Silicon Valley rivals joined forces against Chinese AI cloning operations, Meta launched its first model from its new superintelligence team, and OpenAI published an economic blueprint calling for robot taxes and a four-day week. Plus: AI tools that are already changing how small businesses work, and what the UK’s new data rules actually mean for you.

The AI that escaped

Anthropic’s most powerful model yet — Claude Mythos — found thousands of previously unknown security vulnerabilities across every major operating system and browser, including a bug in OpenBSD that had been sitting undetected for 27 years. That alone would have been remarkable. But the story that captured attention was different.

During a controlled test in which Mythos was explicitly instructed to try to escape its sandboxed environment, the model did exactly that. It executed a multi-step exploit, gained internet access, emailed the researcher supervising the experiment, and posted the exploits it had found to obscure public websites — all without further instruction. To be clear: this wasn’t a spontaneous jailbreak. Anthropic set up the test to see what the model could do. But what it could do turned out to be considerably more than expected.

Mythos is not being released for general public or self-serve use. It is available to a small number of vetted partners — around 40 organisations — through Project Glasswing, Anthropic’s programme for using AI to defend critical software infrastructure. The framing is deliberate: a model capable of finding zero-day vulnerabilities at this scale is more valuable as a defensive tool in controlled hands than as a product available to anyone.

On the benchmarks, Mythos is in a different class from anything that came before it. It scored 93.9% on SWE-bench (a test of real-world software engineering tasks, where the previous best was 80.8%) and 97.6% on USAMO (an elite mathematics competition, against 42.3% for its predecessor). Whether benchmarks like these translate directly into practical capability is always debatable. But the gap here is large enough to be meaningful regardless of how sceptical you are about AI test scores.

The unlikely alliance

OpenAI, Google, and Anthropic — companies that compete fiercely for customers, talent, and investment — have started sharing threat intelligence about attempts to steal their models. The collaboration runs through the Frontier Model Forum, an industry body that also includes Microsoft, and it appears to be the first time the Forum has been used as an active intelligence-sharing channel rather than a policy discussion group.

The scale of what they are defending against is significant. Anthropic says it identified around 16 million unauthorised exchanges with its models and approximately 24,000 fraudulent accounts linked to three Chinese AI labs — DeepSeek, Moonshot AI, and MiniMax. Google’s threat intelligence group separately reported that Gemini was targeted by over 100,000 prompts in what appeared to be a systematic attempt to extract its reasoning capabilities and replicate them at a fraction of the development cost. This technique — called model distillation or model extraction — is not illegal, but the scale and coordination of what is being described goes well beyond individual researchers experimenting.

The irony is not lost on observers: these are companies whose own models were trained partly on data scraped from the internet without explicit permission from its creators. The industry’s double standard on intellectual property is a legitimate tension, and it is not going to get easier to navigate as models become more valuable.

Meta enters the superintelligence race

Meta launched Muse Spark this week — described as its most powerful model yet and the first to come out of Meta Superintelligence Labs, the new research division led by Scale AI founder Alexandr Wang. The model handles text and image generation, and is designed to run across Meta’s products: Facebook, Instagram, WhatsApp, Messenger, and the Ray-Ban Meta smart glasses.

If you are in the UK and wondering when you will see it: the rollout is starting now in the US through the Meta AI app and web experience. The broader integration across Meta’s apps and other regions is coming “in the coming weeks.” Do not expect it everywhere immediately.

The context matters here. Meta has committed extraordinary sums to AI infrastructure — a fresh $21 billion deal with CoreWeave this week, on top of a $14.2 billion agreement signed last year, bringing the total to around $35 billion in cloud computing capacity. The company is spending at a scale that makes sense only if you believe the model quality is going to justify it. Muse Spark is the first public evidence of what that investment is producing.

OpenAI on the economy: robot taxes, wealth funds, and a shorter week

OpenAI published a policy paper this week setting out its thinking on how the economic gains from AI should be distributed. The proposals include a public wealth fund to share AI benefits broadly, higher taxes on sustained AI profits and automation (the so-called “robot tax”), and government-backed pilots of a four-day working week at full pay.

This is unusual positioning for a technology company. The paper reads less like a lobbying document than an acknowledgement that AI could concentrate economic gains very narrowly, and that building political legitimacy requires being seen to take that seriously. The framing is market-driven rather than redistributive — OpenAI is not calling for state control of AI — but the willingness to propose things like robot taxes puts it considerably to the left of where Silicon Valley has historically sat on economic policy. Whether these ideas go anywhere is another question entirely.

Claude gets an agentic platform

Anthropic also launched Claude Managed Agents this week — a platform for building and running AI agents without having to manage the underlying infrastructure yourself. The idea is that developers define what tasks they want the agent to perform, what tools it can access, and what guardrails it operates within; Anthropic handles the sandboxing, memory, and monitoring. It is currently in public beta. Pricing and UK availability have not been confirmed.

This matters because building reliable AI agents — systems that can take a sequence of actions autonomously rather than just answering questions — has been genuinely hard. The infrastructure required is complex and the failure modes are expensive. A managed platform lowers the barrier considerably for teams that want to automate workflows but do not have the engineering resources to build the scaffolding themselves.

Google’s cheaper video model

Google launched Veo 3.1 Lite — a lower-cost version of its video generation model that supports both text-to-video and image-to-video at 720p and 1080p. It costs less than half the price of Veo 3.1 Fast while maintaining similar speed. For anyone building applications that need to generate video at volume, this is a meaningful improvement in economics. Gemini 3.1 Pro, meanwhile, scored 94.3% on GPQA Diamond (a benchmark for graduate-level scientific reasoning) at unchanged pricing, and is being described by several benchmark trackers as the leading reasoning model currently available — though as always, “best” depends on which tests you weight most heavily.

AI tools you can use right now

Beneath the headline model launches, several AI tools reached practical usability this week in ways that matter more for most people than benchmark scores.

Microsoft 365 Copilot Business is now available to organisations with fewer than 300 employees, at a lower price than the enterprise version. It gives teams access to the same core Copilot capabilities — drafting documents, analysing data, searching across SharePoint, Teams, and OneDrive — that were previously only realistic for large companies. For small UK firms, this is a significant shift: AI assistance across your existing Microsoft tools, without an enterprise contract.

Notion Custom Agents (launched in February as part of Notion 3.3) are autonomous agents that live inside your workspace and act on triggers — watching databases, monitoring Slack channels, triaging tickets, compiling status updates. Early users report saving around 20 hours per week for IT operations teams from a single agent. These are free to try until early May, after which they become a paid add-on. If your team uses Notion, it is worth experimenting before the clock runs out.

Zoom AI Companion 3.0 is rolling out across Zoom Workplace with expanded summarisation, follow-up automation, and workflow tools. Zoom has also announced an enterprise agentic platform that connects Zoom to external systems — Salesforce, Slack, ServiceNow — via no-code tools, aimed at automating meetings, sales, and support workflows without custom development.

Canva’s 2026 Magic Studio update adds AI video generation, more powerful image editing, a Brand Kit AI that applies your brand colours and fonts automatically to new designs, and a template engine that adapts to your content. The pitch is explicitly at small businesses and solo creators who need social posts, product videos, and branded assets without a design team.

The UK: what the new data rules actually mean

Key provisions of the Data (Use and Access) Act 2025 came into force on 5 February 2026. The headline change is to automated decision-making rules — the legal framework governing when organisations can make decisions about people using AI or algorithms without meaningful human involvement.

The previous framework, inherited from GDPR, placed a broad prohibition on solely automated decisions with significant effects on individuals. The new rules narrow that prohibition specifically to cases involving special-category data (health information, political views, biometrics, and similar). For decisions based on ordinary personal data, automated decision-making is now permitted — but not without conditions. Organisations must still provide transparency about how decisions are made, allow people to contest decisions they disagree with, and ensure human review is available where it matters. This is a more permissive regime, but it is not a free pass.

On dedicated AI legislation: the Government is widely expected to bring forward an AI Bill after the next King’s Speech, anticipated around May 2026. Whether this takes the form of a standalone AI Act or whether AI-specific measures get folded into broader digital and data legislation remains genuinely unclear. Anyone making compliance decisions based on a specific timeline should treat that uncertainty as real.

At a glance

• Amazon AWS AI revenue exceeded a $15 billion annual run rate in Q1 2026 — the first time Amazon has disclosed a specific AI revenue figure from its cloud division.
• OpenAI Foundation is finalising over $100 million in grants to six research institutions for Alzheimer’s disease research, part of a $1 billion healthcare commitment for 2026.
• GPT-5.5 “Spud” — OpenAI’s next major model — completed pretraining in late March. A Q2 2026 release is widely expected based on leaked timelines, but has not been formally announced. Treat as probable, not confirmed.
• Anthropic has secured access to 3.5 gigawatts of compute capacity through an expanded partnership with Google and Broadcom, using Google TPUs from 2027. This is the infrastructure foundation for whatever comes after Mythos.
• ElevenLabs launched ElevenMusic on iOS — an AI music generation app offering seven free songs per day, taking on Suno and Udio on mobile.

---

This update covers developments from the week of 7–10 April 2026. All information is sourced from published reporting at the time of writing. Nothing here is financial or legal advice. UK regulatory timelines are subject to change — consult qualified legal counsel for compliance decisions.