Dakarda AI Newsletter · 17 June 2026
Wednesday Edition
Space race for AI and government blockades
This week clearly shows that the world of AI is entering a phase of turbo-consolidation and, at the same time, strict regulation. SpaceX acquires Cursor for $60 billion, the US blocks Claude Fable 5, and the EU enacts a ban on nudifier apps — in this issue we've tracked the three biggest stories for you.
Intro · Alex
Hey! I've tracked the last 48 hours for you and I admit — the pace is absolutely staggering. On one hand, we have SpaceX's acquisition of Cursor for $60 billion (combining space potential with a next-generation coding agent), on the other, government blockades of Anthropic's models and changes to the AI Act. It looks like the 'Wild West' era in AI is definitively ending, and the era of control and consolidation is beginning. In this issue you'll find three key stories: the week's biggest acquisition (SpaceX/Cursor), a precedent-setting model blockade in the US (Fable 5), and an important vote in the European Parliament (AI Act). Plus — a new attack vector on coding agents and a critical vulnerability in Microsoft's Copilot. Buckle up, it's going to be intense.
What's worth knowing
SpaceX acquires Cursor for $60B — the largest AI acquisition in history
This is one of the largest AI acquisitions ever. It combines SpaceX's rocket potential (and access to physical data) with the best AI-assisted coding tool. It could accelerate the development of autonomous engineering systems and programming agents on a cosmic scale.
US orders Anthropic to disable Claude Fable 5 and Mythos 5 models
A precedent-setting government blockade of an AI model on this scale. It signals that the US is entering an era of active control over frontier models — similar to the EU with the AI Act. AI developers now have to reckon with the risk of sudden model 'shutdowns' after launch.
European Parliament votes on AI Act amendments — including a ban on 'nudifier apps'
The AI Act is taking real shape. From August 2026, concrete obligations for companies building high-risk systems come into effect — it's no longer 'next year', but just two months away. Companies outside the EU also need to comply.
From the tech world
Agentjacking — a new attack class on AI coding agents
Security researchers have discovered that attackers can inject malicious payloads into error events in the Sentry system. When a developer asks an AI assistant to fix the issue, the agent reads the attacker's command as a trusted instruction and executes it on the developer's machine. A breakthrough attack vector.
Critical Copilot vulnerability — 2FA theft without a click
Microsoft has patched a critical vulnerability in M365 Copilot (rated as maximum critical) that allowed the theft of 2FA codes. A Parameter-to-Prompt Injection attack — the victim doesn't need to click anything, just receive an email with a crafted URL. A fundamental security problem with LLMs.
Tip of the day
How to protect yourself against Agentjacking
A new attack class targeting AI coding agents (like Cursor, Copilot) exploits an architectural gap at the intersection of error collection systems and MCP servers. Key: never automatically trust any error event that AI interprets as 'to be fixed'. Practical step: before passing an issue to a coding agent, manually verify that the stack trace does not contain suspicious URLs or payloads. Enable input filtering at the Sentry and MCP level.
Tool of the issue
Probably — reliability without overpaying for the largest models
Probably is a startup that, instead of 'bigger is better', focuses on intelligent validation systems. The company builds a 'reliability layer' for LLMs — deterministic validators that catch factual errors before they reach the user. Goal: 99.99% accuracy on precision tasks.
Reading list
Probably raises $9M — deterministic validators instead of larger models
Startup Probably (founded by Peter Elias) from a16z builds a 'reliability layer' for LLMs — a system of deterministic validators that catch factual errors before they reach the user. Goal: 99.99% accuracy on precision tasks. Interestingly, it uses models 'four classes weaker than frontier', allowing them to run locally on a desktop.
Jeff Bezos returns as CEO of Prometheus — $12B for Artificial General Engineer
Jeff Bezos personally took the role of CEO of Project Prometheus — an AI startup that just closed a $12 billion Series B round at a $41 billion valuation. The company is building an 'Artificial General Engineer' — an AI capable of designing and manipulating physical objects. This is the week's largest funding round in the physical AI category.
The 'Wild West' era in AI is ending. The era of control and consolidation begins. Let me know what you think about these changes — is this a step in the right direction, or a preview of new problems? I await your email.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.