22 April 2026: Google’s Agent Platform and OpenAI’s Image Upgrade Lead a Packed Day

Google’s Cloud Next event brought dual AI chips and a fresh enterprise agent platform, OpenAI launched a smarter image generator, and Anthropic found itself investigating unauthorised access to a powerful new cybersecurity model. Meanwhile, SpaceX quietly locked in a $60 billion option to buy AI coding startup Cursor, signalling just how high the stakes in AI tooling have become.

Google split its new TPU into two separate chips at Cloud Next ’26, making a pointed move against Nvidia’s data centre dominance. The eighth generation of Google’s Tensor Processing Unit arrives as two distinct products: TPU 8t for training large models, and TPU 8i for inference, the process of running a trained model to generate responses. Separating the two workloads allows Google to optimise each chip independently, potentially offering better performance per cost than a one-size-fits-all design. Alongside the chip announcement, Google launched its Gemini Enterprise Agent Platform, a revamped developer environment built on Vertex AI that manages the full lifecycle of AI agent fleets. The platform is designed to reduce what the industry calls “agent sprawl,” the proliferation of disconnected AI agents that large organisations have struggled to govern. Google also unveiled Workspace Intelligence, an AI layer woven across Google Docs, Sheets, and Meet that understands relationships between data stored across different Workspace apps. Google announced a $750 million fund to help consulting partners adopt and deploy these tools for enterprise clients.

OpenAI launched ChatGPT Images 2.0 today, giving the tool what the company calls “thinking capabilities” and the ability to search the web before generating an image. The update means the system can pull in current information, diagrams, or reference material before producing output, rather than relying solely on what it was trained on. Users can now generate complex infographics, slides, maps, detailed charts, and even manga panels from a single prompt, with substantially more reliable text rendering than previous versions. The upgrade is available to ChatGPT users now and represents a significant step forward for anyone producing visual content professionally. Small businesses that currently rely on stock imagery or simple design tools should take note: the tool can now handle layouts that previously required a professional graphic designer or a paid software subscription.

Anthropic is investigating an incident in which an unauthorised group appears to have gained access to Mythos, its restricted new AI model built for cybersecurity research. Mythos is not a publicly available product. It has been shared with a small number of government agencies and researchers due to its ability to find, and potentially exploit, software vulnerabilities at scale. The breach, first reported by TechCrunch and confirmed by Anthropic as an active investigation, raises questions about how even tightly controlled AI tools can slip beyond their intended boundaries. In a separate but related development, Mozilla published results from an internal test using Mythos that uncovered 271 vulnerabilities in Firefox. Mozilla said the exercise demonstrated that AI tools could shift the balance of cybersecurity toward defenders, allowing security teams to find and patch flaws faster than attackers can exploit them. Anthropic’s position is unusual: its most capable new model is simultaneously generating serious safety concerns and demonstrating genuine defensive utility.

SpaceX has secured the option to acquire Cursor, one of the leading AI coding assistants, for up to $60 billion. Cursor, developed by Anysphere, has become a widely used tool for software developers, allowing them to write, edit, and debug code using natural language. The deal gives SpaceX the right, but not the obligation, to complete the acquisition later this year. Reports suggest the arrangement reflects a deliberate effort to fill a gap in xAI’s capabilities, as Cursor sits alongside rivals such as GitHub Copilot and Anthropic’s Claude Code. The valuation underlines how seriously technology companies now treat AI coding tools, which have moved from novelty to core development infrastructure at many firms in fewer than three years.

Sullivan and Cromwell, one of Wall Street’s most prominent law firms, has apologised to a US federal bankruptcy court after discovering that a major filing contained multiple AI hallucinations. Hallucinations are instances where an AI system produces plausible-sounding but false information, including invented case citations, misquoted rulings, or fabricated legal arguments. The firm said the errors were introduced during drafting and were not caught before submission. The incident joins a growing list of cases in which professional firms have faced embarrassment after relying on AI-generated output without sufficient human review. Regulators and bar associations in multiple jurisdictions are now moving to formalise requirements around AI disclosure in legal filings, and this case is likely to accelerate those efforts.

Sony AI demonstrated what it describes as the first robot to achieve expert-level performance in a physical sport, with its autonomous table tennis system beating top human players in a controlled competition. The robot uses what Sony calls “agentic AI” to read an opponent’s movements, predict ball trajectory, and execute return shots in real time. Achieving consistent expert performance in an uncontrolled physical environment is considerably harder than in simulation, as the system must account for real-world variance in lighting, bounce, and spin. While the application may seem narrow, the underlying technology, an AI that perceives, decides, and acts with physical precision under pressure, has clear implications for robotics in manufacturing, logistics, and assistive care.

Tencent and Alibaba are in advanced talks to invest in DeepSeek, the Chinese AI research lab, at a valuation exceeding $20 billion. DeepSeek drew significant international attention earlier in 2026 after publishing research suggesting it had trained highly capable models at a fraction of the cost of comparable Western systems. A valuation of $20 billion would position it among the most valuable AI research organisations outside the United States. The involvement of Tencent and Alibaba could give the lab access to significant distribution infrastructure within China, and Western analysts are watching the deal closely for any implications around DeepSeek’s future publication practices.

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

• Mira Murati’s Thinking Machines Lab signed a multi-billion dollar deal with Google Cloud to access AI systems built on Nvidia’s GB300 chips. TechCrunch
• Tencent’s QClaw, an AI agent app based on OpenClaw, opened its international beta for Windows and macOS. KrASIA
• Vast Data, which builds software infrastructure for managing large datasets for AI workloads, raised $1 billion in a Series F at a $30 billion valuation, backed by Nvidia. CNBC
• Apple named hardware chief John Ternus as its new CEO, with Tim Cook moving to executive chairman. Bloomberg
• X (Twitter) rolled out more than 75 AI-powered custom timelines powered by Grok for Premium subscribers. MacRumors
• Anthropic is reported to be testing whether to remove Claude Code access from its $20 per month Pro subscription plan. PCWorld
• The UK’s High Court ruled that London’s Metropolitan Police may continue using live facial recognition technology. The Register
• SK Hynix announced plans to invest $12.85 billion in a new advanced packaging facility in South Korea to meet AI demand for high-bandwidth memory. Reuters

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 evening.