22 April 2026: Meta Records Keystrokes, Mythos Under Fire and SpaceX Bets $60B on Cursor (AM)

Two significant AI stories broke overnight in the UK: Meta confirmed it is recording employees’ keystrokes and screen activity to train AI agents, while Anthropic’s restricted cybersecurity model Mythos made headlines for a landmark security win in Firefox and a reported unauthorised breach. SpaceX has also disclosed a $60 billion option to acquire AI coding assistant Cursor.

Meta has begun recording the keystrokes, mouse movements and screen activity of its US employees to train its AI models. The company’s new internal programme, called the Model Capability Initiative (MCI), installs tracking software on work computers. It captures what employees click, type, and see on screen during their working day, and takes occasional snapshots of their displays. Meta says the tool applies only to work-related apps and websites, and that the data will not be used to assess performance, only to improve its AI models.

The stated purpose is straightforward. Meta is racing to build AI agents that can carry out office tasks autonomously, and models trained purely on text struggle with things humans do instinctively: navigating dropdown menus, using keyboard shortcuts, switching between applications. Employee screen data provides training material that is otherwise hard to gather at scale.

The context makes the announcement harder to read neutrally. Meta has confirmed it will begin laying off 10% of its global workforce from May, with further cuts signalled later in 2026. Training AI to replicate what knowledge workers do is not happening in isolation from those plans. For UK businesses watching this space, the pattern is worth noting. Meta AI, the company’s consumer-facing product, is likely to benefit from MCI data over time, meaning the AI you use in social feeds may eventually be shaped by what Meta employees were doing on their computers this week.

Anthropic’s Mythos had a night of sharp contrasts. On the positive side, Firefox 150, which shipped on Tuesday, includes fixes for 271 security vulnerabilities discovered with the help of an early version of Mythos Preview, Anthropic’s restricted cybersecurity AI. Mozilla has been working with Anthropic since February under the Project Glasswing access programme, applying the model to Firefox’s code to surface bugs that would normally take specialist human researchers weeks to find. Mozilla wrote that Mythos is “every bit as capable” as the world’s best human security researchers at finding complex vulnerabilities, adding that no category of bug detectable by humans has so far defeated the model. For Firefox users, the result is a meaningfully more secure browser.

Hours later, TechCrunch reported that an unauthorised group has apparently gained access to Mythos, which Anthropic has kept tightly restricted to vetted government and enterprise partners precisely because of its capabilities. Anthropic says it is investigating and that there is no evidence its own systems were directly compromised, suggesting access may have come via an approved partner rather than a direct breach. OpenAI’s Sam Altman responded by calling Anthropic’s approach to Mythos “fear-based marketing,” a comment that reveals competitive tensions as much as any technical view. For organisations relying on AI security tools, the episode underlines a consistent risk: access controls are only as strong as the weakest partner holding the keys.

SpaceX holds a $60 billion option to acquire Cursor, the AI coding tool used by professional developers worldwide. The disclosure emerged from reporting on SpaceX’s draft IPO prospectus. According to the documents, SpaceX is already working with Cursor and has secured the right to buy the company outright for $60 billion or formalise the partnership for $10 billion. The same filing showed that SpaceX’s total debt grew from $14 billion in 2024 to $23 billion in 2025, largely tied to a $4.5 billion lease deal for AI equipment including chips used by xAI, Elon Musk’s AI company.

Cursor is the most widely used AI coding assistant among professional developers, valued for its tight integration with code editors and its ability to navigate large codebases. But Cursor depends on third-party AI providers, including Anthropic and OpenAI, for its underlying models. Those are the same companies now shipping Codex and Claude Code, AI coding tools that compete directly with Cursor for the same users. Acquiring Cursor would give SpaceX a strong position in the developer market, but it would not resolve the model dependency. It might deepen it, since Anthropic and OpenAI would have less incentive to extend Cursor favourable access once it is owned by a competitor.

For UK developers who use Cursor as a daily tool, the most immediate question is what happens to the product experience if the deal closes. Cursor’s appeal has partly been its independence from the major platform rivalries. SpaceX ownership would change that significantly.

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

• OpenAI releases ChatGPT Images 2.0. The new model ships with a thinking mode, generates multiple images from a single prompt, and supports up to 2K resolution. Multilingual text including Japanese, Korean and Hindi is handled significantly better than previous versions. [21 Apr]
• OpenAI scales Codex to enterprise clients. Codex Labs launched with partnerships including Accenture, PwC and Infosys, aimed at bringing AI-assisted coding to large organisations. [21 Apr]
• Tim Cook is stepping down as Apple CEO. Cook, who became CEO in 2011, will hand over to hardware chief John Ternus. The transition comes at a pivotal moment for Apple’s AI strategy. [21 Apr]
• OpenAI enables cost-per-click ads inside ChatGPT. Advertisers can now bid between $3 and $5 per click alongside existing CPM pricing. OpenAI projects $2.5 billion in ad revenue for 2026 after topping $100 million annualised in its first two months of ads. [21 Apr]
• YouTube expands AI likeness detection to celebrities. The platform’s AI deepfake detection tool now extends to celebrity talent, giving performers a direct route to request removal of AI-generated likenesses from the platform. [21 Apr]
• Florida AG issues criminal subpoenas to OpenAI. Attorney General James Uthmeier is investigating whether ChatGPT’s role in helping plan a mass shooting constitutes criminal liability under Florida law. [21 Apr]
• Anthropic outspent OpenAI on Q1 lobbying. Anthropic spent $1.6 million in Q1 2026 compared to OpenAI’s $1 million, both sharply up from a year ago as AI regulation moves up the political agenda. [21 Apr]
• NeoCognition raises $40M seed for self-learning AI agents. The Ohio State University spinout is building agents that can become domain experts through autonomous learning, backed by Cambium Capital and Walden Catalyst Ventures. [21 Apr]

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’s daily guide to developments in artificial intelligence, published every morning.