21 April 2026: The AI That Can Hack Your Browser, Ads Inside ChatGPT, and Half a Billion for UK AI

Two days in AI that will be worth remembering. Anthropic announced a model so capable at finding software vulnerabilities that the company decided the public should not have access to it. OpenAI opened ChatGPT to cost-per-click advertising. And the UK government began handing out contracts from a \u00a3500 million sovereign AI fund. Here is what it all means.

Anthropic has built an AI that can independently find and exploit unknown software vulnerabilities, and it has decided not to release it to the public. The model, Claude Mythos Preview, was announced on 20 April after internal testing showed it produced working exploits in Firefox security flaws 181 times where its predecessor, Opus 4.6, managed it twice. It independently discovered and exploited zero-day vulnerabilities, meaning previously unknown weaknesses, across every major operating system and web browser in the test environment.

Instead of a standard release, Anthropic has made Mythos Preview available only to a selected consortium of technology companies through a programme called Project Glasswing. The stated purpose is defensive: use the model to identify and fix critical software weaknesses before adversaries with similar capabilities can exploit them. A companion model, Opus 4.7, has been released publicly, with cybersecurity capabilities deliberately reduced during training and automatic filters that block high-risk hacking requests built in from the outset.

For most people, nothing changes today in practical terms. For organisations running their own software infrastructure, the signal is sharper. The AI now available to well-resourced attackers is operating at a level that makes faster patching and vulnerability scanning a genuine priority. Anthropic also released a reliability and speed update for Claude Code, its coding assistant, covering faster session resumption, quicker tool startup, and smarter slash-command search.

OpenAI has switched on cost-per-click advertising inside ChatGPT, making it possible for businesses to buy paid placement within the world\u2019s most widely used AI assistant. The change, confirmed on 21 April, lets advertisers bid between roughly \u00a32.40 and \u00a34 per click using an ad management platform that OpenAI now operates directly. This marks a shift away from the impression-based model launched around ten weeks ago, where advertisers paid a flat rate per thousand users reached regardless of engagement.

For small businesses, this matters in a direct way. Millions of people use ChatGPT to research purchases, compare services, and draft buying decisions. Appearing in those conversations is a different proposition from appearing in a search results page, and CPC pricing means businesses can enter with a controlled budget rather than the large upfront minimums that characterised the launch phase. OpenAI has projected \u00a32 billion in advertising revenue for 2026, and the pilot had already passed \u00a380 million in annualised revenue in under two months.

The UK government has opened the first procurement round on a \u00a3500 million Sovereign AI fund, with the initial tranche worth \u00a380 million and applications expected by mid-May. The fund is being delivered by the Department for Science, Innovation and Technology and chaired by James Wise of Balderton Capital. What makes the terms notable is the intellectual property arrangement: successful bidders will retain ownership of the IP they build, with the government keeping usage rights. That is a deliberate departure from standard central-government contracting, where the Crown typically claims ownership of commissioned work.

Challenge areas map to the government\u2019s AI Opportunities Action Plan: scientific discovery, health and social care, national security and defence, transport, and energy. First-round bidders must make contact by 16 May, with the main competition running in July and delivery timelines of 12 to 24 months. For UK AI companies and university research groups, this is an open bidding process with genuinely unusual terms. Whether the fund reaches innovative teams rather than established contractors will be the real test.

MIT Technology Review published its first annual list of 10 Things That Matter in AI Right Now on 21 April, launching it at EmTech AI 2026, which opened today on the MIT campus in Cambridge, Massachusetts. The list was compiled through internal voting by the publication\u2019s AI reporting team and identifies the areas they intend to follow most closely across the rest of 2026. Among the ten: AI companions, mechanistic interpretability, generative coding, and hyperscale data centres.

Mechanistic interpretability is the research field attempting to explain what is actually happening inside large language models, including why they sometimes present inaccurate information with apparent confidence. If that work matures, the implications are real: organisations relying on AI outputs without human review would have a more principled basis for deciding when that is safe. The conference runs until 23 April and this year\u2019s theme is the transition of AI from experimentation into core business infrastructure, which is where most organisations actually are right now.

Here is everything else worth knowing from today\u2019s AI news.

• ChatGPT 90-minute outage: ChatGPT experienced a partial outage on 20 April affecting over 8,700 UK users at peak. Login, voice mode, image generation, and the API were all disrupted before OpenAI deployed a fix and declared recovery. [20 Apr]
• OpenAI Spud model approaching: OpenAI\u2019s next model, codenamed Spud and expected to launch as GPT-5.5 or GPT-6, completed pretraining on 24 March. Sam Altman described the release as a few weeks away. It was trained using more than 100,000 H100 GPUs at the Stargate facility in Texas. [April 2026]
• Anthropic adds ID verification for some users: Anthropic has added government ID and selfie verification for select Claude users, processed through a service called Persona. Accepted documents include passports and driving licences. The data will not be used to train models, Anthropic says. [14 Apr]
• Google Gemma 4 released open source: Google released Gemma 4 under the Apache 2.0 licence on 2 April in four model sizes. All variants are designed to run fully offline on edge devices including smartphones and Raspberry Pi hardware, with context windows up to 256,000 tokens. [2 Apr]
• NVIDIA Ising for quantum computing: NVIDIA launched Ising on 14 April, a family of open AI models for quantum processor calibration and error correction. Early adopters include Harvard, Lawrence Berkeley National Laboratory, and the UK\u2019s National Physical Laboratory. [14 Apr]
• YouTube extends deepfake detection: YouTube is expanding its AI-powered deepfake detection tool to celebrities and their representatives, giving them a way to find and request removal of AI-generated likeness content from the platform. [April 2026]

This is a 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\u2019s daily guide to developments in artificial intelligence, published every evening.